FUSED-Net: Detecting traffic signs with limited data

Md Atiqur Rahman; Nahian Ibn Asad; Md Mushfiqul Haque; Md Bakhtiar Hasan; Sabbir Ahmed; Md Hasanul Kabir

doi:10.1371/journal.pone.0359293

. 2026 Oct 7;21(10):e0359293. doi: 10.1371/journal.pone.0359293

FUSED-Net: Detecting traffic signs with limited data

Md Atiqur Rahman ^1,^#, Nahian Ibn Asad ^1,^2,^#, Md Mushfiqul Haque ^1,^3,^#, Md Bakhtiar Hasan ¹, Sabbir Ahmed ¹, Md Hasanul Kabir ^1,^*

Editor: Baicang Guo⁴

PMCID: PMC13645229 PMID: 42842623

Abstract

Automatic Traffic Sign Recognition is paramount in modern transportation systems. However, curating large-scale datasets for diverse traffic sign detection remains impractical. In this context, we present FUSED-Net, a novel approach that enhances Few-Shot Object Detection (FSOD) for traffic signs using limited data. FUSED-Net integrates Faster RCNN with Unfrozen Parameters, Pseudo-Support Sets, Embedding Normalization, and Domain Adaptation to improve detection accuracy. Unlike conventional methods, FUSED-Net keeps all parameters unfrozen during training, enabling it to learn effectively from limited samples. A Pseudo-Support Set is generated through data augmentation, enhancing performance by compensating for the scarcity of target domain data. Embedding Normalization reduces intra-class variance, standardizing feature representations. Domain Adaptation, achieved by pre-training on a diverse traffic sign dataset, improves model generalization. Experimental results on the BDTSD dataset demonstrate that FUSED-Net achieves $2.4 \times$ , $2.2 \times$ , $1.5 \times$ , and $1.3 \times$ improvements in mAP under 1-shot, 3-shot, 5-shot, and 10-shot scenarios, respectively, compared to state-of-the-art FSOD models. Additionally, FUSED-Net achieves superior performance on the cross-domain FSOD benchmark across multiple settings. The source code and the URLs to download the datasets are available at https://github.com/180041123-Atiq/FUSED-Net.

Introduction

Traffic sign recognition plays a crucial role in enhancing road safety, supporting autonomous vehicle technology, and assisting in efficient driving [1]. However, interpreting traffic signs in real-time can be challenging because of issues such as environmental factors, distance, and limited visibility [2], which can lead to mis-recognition and potentially catastrophic accidents [3]. To mitigate these risks, Advanced Driver Assistance Systems [4] and Automated Driving Systems [5] have been developed in the literature, with traffic sign detection being a key component of their functionality. Hence, accurate traffic sign detection is vital for the effective operation of these systems, especially in real-world conditions. In addition to conventional environmental challenges, the robustness and security of traffic sign perception systems against adversarial attacks have also received increasing attention. Recent studies have investigated benchmark frameworks for meta-adversarial defense [6], robust pre-training and fine-tuning strategies for traffic sign classification in autonomous driving [7], and self-adaptive defense mechanisms for vision-language models against full-dimension attacks in autonomous driving [8]. These studies highlight the importance of developing reliable and robust perception systems for safety-critical autonomous driving applications.

Over the years, traffic sign detection systems have evolved from manually engineered features to Deep Learning (DL)-based methods, particularly Convolutional Neural Networks (CNNs) [9–13]. These methods are generally categorized into two categories: one- and two-stage detectors [14]. Recent advances in object detection have further explored geometry-guided representations [15], structure-aware transformer architectures [16], and multi-granularity prompt learning [17], demonstrating the growing diversity of strategies for improving object localization and recognition under challenging conditions. One-stage detectors such as YOLO [18,19] and DETR [20] treat sign detection as a regression problem. Here, a single convolutional network simultaneously predicts multiple bounding boxes around the traffic sign along with the class probabilities for those boxes, offering faster processing but less accuracy [21,22]. In contrast, two-stage detectors, such as Faster R-CNN [23], first generate region proposals around the signs and then classify them, leading to higher accuracy but slower inference.

Despite the emergence of DL-based methods that have introduced significant progress in traffic sign detection [24–30], these methods require extensive annotated training data to achieve optimal performance. However, publicly available datasets, such as TT100K [24], primarily represent traffic signs from specific regions like China, which are significantly different from those in underdeveloped or developing countries, such as Bangladesh. In these countries, the condition of traffic signs is often compromised due to limited maintenance, occlusion, or wear and tear, introducing additional challenges that are not well-represented in such datasets. Consequently, a model trained on one country’s dataset(s) might not perform well on another without being fine-tuned on a sufficient number of target samples. In this context, collecting large-scale annotated datasets for traffic signs across every country is impractical due to the diversity of sign designs and environmental conditions.

Even with the advancements in hardware technology and the generation of massive datasets that have propelled the DL-based approaches towards human-like capabilities, they are still unable to learn effectively from only a few samples [31]. Unlike humans, who can quickly adapt to new tasks using prior knowledge, DL models require extensive data to perform effectively [32]. This disparity in learning efficiency is particularly challenging in the context of traffic sign detection in underrepresented regions, such as Bangladesh, where annotated samples are scarce. In this connection, the recent family of Few-Shot Learning (FSL)-based models can play an instrumental role with their remarkable ability to learn with limited data and have already proved their worth in a diversified arena of computer vision tasks [33], making them an ideal choice for this problem domain.

Recent research endeavors have applied FSL-based approaches to problems such as object detection and classification, collectively referred to as Few-Shot Object Detection (FSOD) [34]. FSOD is especially useful in situations where it is difficult to obtain large labeled datasets, such as with traffic signs from remote locations that are underrepresented due to negligence and lack of preservation efforts [35]. Hence, in this work, we propose FUSED-Net, a fusion of Faster RCNN with Unfrozen Parameters, Pseudo-Support Sets, Embedding Normalization, and Domain Adaptation while reducing data requirements. Empirical analysis shows that our proposed model FUSED-Net achieves superior performance compared to the state-of-the-art FSOD architecture on the target traffic sign dataset with limited data, significantly reducing the need for large datasets. This advancement can greatly improve Traffic Sign Detection and Recognition Systems, allowing them to identify traffic signs accurately in various real-world conditions, tackling the challenges of data scarcity. Our contributions are:

Unlike conventional few-shot object detection models that selectively freeze parameters during training, we propose an architecture where the entire network remains unfrozen throughout the training process. This approach allows the model to fully adapt to the limited number of available samples, enabling more robust learning and representation extraction from the few examples in the target domain.
To mitigate the challenge of data scarcity inherent in few-shot learning, we leverage a novel technique called pseudo-support sets that are generated by applying data augmentation techniques to the few available labeled samples per class. This approach not only increases the diversity of the training data but also significantly enhances the model’s detection performance by compensating for the limited sample size in the target domain.
We incorporate a cosine similarity-based classifier in our architecture to implement Embedding Normalization (EN). By reducing intra-class variance, EN ensures that the feature representations of traffic signs are more standardized and consistent across the limited number of samples. This improves the model’s ability to differentiate between subtle class differences, leading to more precise detection and classification.
We enhance the generalization capability of the proposed architecture through a Domain Adaptation strategy. By pre-training the model on a diverse traffic sign dataset that is distinct from the target query set, we enable the model to better adapt to the specific characteristics of the target domain. This approach significantly boosts performance on datasets with minimal overlap between the training and test domains, proving crucial for cross-domain few-shot detection tasks.

The remainder of this paper is organized as follows. In the Related works section, we review the relevant literature, focusing on existing approaches to Few-Shot Traffic Sign Detection and Few-Shot Object Detection in general, highlighting the current challenges and limitations. The Proposed methods section presents our proposed FUSED-Net, detailing the baseline and the modifications made to the Faster R-CNN framework. In the Results and Discussion section, we present the outcome of extensive experiments on multiple datasets demonstrating the performance of FUSED-Net in various few-shot scenarios and compare the results with the state-of-the-art FSOD models. We also discuss the effectiveness of our key contributions and cross-domain generalization performance. The Conclusion section concludes the paper, summarizing our findings and offering directions for future research in few-shot traffic sign detection.

Related works

Few-Shot Object Detection (FSOD) has emerged as a critical research area due to the growing need for models that can effectively detect objects with minimal labeled data [32]. As traditional object detection methods typically require large-scale annotated datasets to achieve high performance, FSOD seeks to overcome this limitation by enabling models to generalize well even when only a few instances of the target objects are available.

The literature on FSOD is extensive and diverse, with several approaches being proposed and refined over the years. According to [36], FSOD techniques can be broadly categorized into three primary categories: transfer learning-based, meta-learning-based, and metric learning-based approaches. Transfer learning in FSOD leverages a pre-trained network’s feature extraction capabilities, which are fine-tuned on a smaller target dataset to adapt its features to a new domain with minimal additional data [37]. This method is powerful because it applies rich features learned from large datasets to specific tasks with limited data. However, ensuring effective generalization to new classes without overfitting remains a challenge. Regardless, simple fine-tuning has sometimes outperformed complex methods like meta-learning, indicating their robustness.

Meta-learning, commonly known as, “learning to learn”, enables rapid adaptation to new detection tasks by training models on diverse range of object classes. However, as highlighted by [37], these methods often require complex architectures and extensive training. At the same time, their effectiveness can be limited by inconsistent evaluation standards in the field, making them less practical for real-world FSOD applications. Metric learning, focused on learning a distance function to differentiate object classes, excels in FSOD by allowing efficient comparison of new instances with few labeled examples. Regardless of their simplicity and efficiency, a fundamental issue with metric learning is that they tend to perform poorly when there are slight visible changes in the support set or query set [38].

Despite the progress in FSOD, there remains a notable gap in the literature regarding few-shot traffic sign detection. Most research in this domain has focused on traffic sign classification [39,40], with very few studies addressing the unique challenges of traffic sign detection. Notable exceptions include [41] and [42], who explored FSOD in traffic scenarios using YOLO-based and transformer-based architectures, respectively. However, these approaches have generally underperformed compared to Faster R-CNN-based architectures, as discussed by [36] and [43]. YOLO-based models, while faster, often struggle with localization accuracy, which is critical in detecting small and closely spaced objects like traffic signs and the performance is even more degraded when the data is scarce [11]. Transformers, on the other hand, incur a high computational cost, making them less practical for real-time detection tasks in traffic environments, and they also pose the requirement of abundant data for high performance. These limitations have led us to adopt Faster R-CNN as our base detector, given its superior performance in terms of localization and its ability to be effectively adapted for few-shot learning through transfer and metric learning techniques [23,37,44–46].

Among the approaches built upon Faster R-CNN, TFA w/cos [37] represents a significant advancement in FSOD by introducing a two-stage detector with a frozen backbone and a cosine similarity-based classifier. This approach was designed to prevent overfitting to the small target dataset by keeping most of the network’s parameters fixed during fine-tuning. However, subsequent studies, such as FSCE [47], have highlighted the limitations of this approach, particularly the potential performance degradation caused by freezing key modules like the Region Proposal Network (RPN). FSCE addressed these issues by introducing a Contrastive Proposal Encoding Loss (CPE Loss) that enhances the model’s ability to differentiate between similar and dissimilar proposals, thus improving intra-class compactness and inter-class variance. However, this architecture still faced challenges in generalizing to novel classes, which is a critical aspect of FSOD.

The decision to freeze certain modules in FSOD architectures has been a subject of debate in the literature. Cross-Domain Few-Shot Object Detection (CD-FSOD) [46], for instance, argues that unfrozen parameters are crucial for enhancing the performance of FSOD models, particularly in cross-domain scenarios. The authors showed that architectures with unfrozen parameters consistently outperformed those with frozen modules, suggesting that the flexibility to update the parameters is key to adapting to new and diverse object categories. As shown in Fig 1, unfrozen parameters can significantly enhance performance, particularly in the context of few-shot traffic sign detection, where the diversity of object instances is relatively limited compared to broader object detection tasks.

Fig 1 — The graph compares the Mean Average Precision (mAP) between frozen and unfrozen training conditions across different state-of-the-art architectures on the Cross-domain FSOD benchmark. The results demonstrate that keeping parameters unfrozen during training consistently enhances performance, as indicated by the higher mAP scores in the unfrozen condition. (Data adapted from [46]).

One key takeaway from TFA w/cos approach is the utilization of a cosine similarity-based classifier to normalize and compare instance-level embeddings, enhancing the model’s ability to capture relationships between instances and classes. As recommended by [37], this approach ensures comparability across different instances, improving generalization and stability in dynamic few-shot learning scenarios where the model must adapt to new classes without forgetting previous ones [48]. This led us to incorporate a cosine-similarity-based classifier in our work to improve performance in data-scarce environments by leveraging this effective embedding normalization.

Again, as per the observations of [44,46], domain adaptation and data augmentation are crucial strategies for improving the generalization of FSOD models, particularly in cross-domain settings where the target categories differ significantly from those in the training data. The CD-FSOD benchmark exemplifies this approach by incorporating a diverse set of categories from different domains, thereby challenging the models to adapt to a broader range of object classes. The CD-FSOD model also employed data augmentation techniques, including both strong and weak augmentations, to enhance the robustness of their models. These techniques, such as color jittering, grayscale conversion, gaussian blur, and cutout patches, help prevent overfitting by introducing variability into the training data. However, the complexity of the CD-FSOD approach, particularly the use of a teacher-student distillation framework and Exponential Moving Average (EMA) updates, may be unnecessary for tasks like traffic sign detection, where the objects have relatively consistent shapes and colors. Our analysis showed that a simpler FSOD model, based on a straightforward Faster R-CNN fine-tuning (FRCN-ft) approach, retained a significant portion of CD-FSOD’s performance while being easier to implement and more computationally efficient. This inspired us to adopt a simpler methodology that utilizes an augmented support set, named pseudo-support set to improve detection performance in our target domain.

The use of attention mechanisms in FSOD has also gained traction, particularly in the work by [44], who proposed an Attention-based Region Proposal Network (A-RPN) utilizing information from the support set images that enables them to focus on the relevant portion of the query set images during inference. This approach aims to improve the model’s ability to detect novel categories by directing attention to relevant regions in the query images. However, our evaluation of A-RPN revealed its poor performance in few-shot traffic sign detection settings. Upon further investigation, we identified a significant disparity in the quality of support images generated by the attention RPN when fine-tuning with the Bangladeshi Traffic Sign Detection Dataset (BDTSD) [35]. This indicates that the effectiveness of attention mechanisms in FSOD highly depends on the quality of the support images, which may not always be guaranteed in real-world scenarios. Consequently, while attention mechanisms hold promise, we chose not to incorporate them into FUSED-Net due to the performance inconsistency observed in our evaluations.

Decoupled Faster R-CNN (DeFRCN) [45] represents another significant contribution to the FSOD field by improving the standard Faster R-CNN architecture to make it more suitable for few-shot learning tasks. DeFRCN introduced multi-stage decoupling and enhanced translation invariant features via Gradient Decoupled Layer (GDL) and Proto Calibration Block (PCB), respectively. Despite these advancements, DeFRCN’s reliance on frozen modules and its inability to enhance classification scores with low-quality region proposals limit its applicability in scenarios where high-quality proposals are not guaranteed. This underscores the importance of robust proposal generation in FSOD, particularly for tasks like traffic sign detection, where the quality of the region proposals can significantly impact the model’s performance.

In conclusion, the extensive body of research on FSOD highlights the diversity of approaches and the ongoing debates regarding the most effective strategies for few-shot learning. Despite the progress made so far, significant challenges remain, particularly in the application of FSOD to specific domains like traffic sign detection. Drawing inspiration from the strengths and limitations of the existing literature, we propose FUSED-Net, which integrates effective strategies from state-of-the-art architectures while introducing novel enhancements. Our methodology leverages the robustness of Faster R-CNN, the adaptability of transfer and metric learning, and the practical benefits of domain adaptation and pseudo-support sets, all tailored to address the unique challenges of few-shot traffic sign detection. The following sections provide a detailed exploration of our proposed approach, demonstrating how FUSED-Net advances the state-of-the-art in this specialized field.

Proposed methodology

Our proposed framework is built upon Faster R-CNN, pre-trained on the MS COCO dataset [49], due to its established robustness in object detection tasks [50]. Recognizing the challenges inherent in few-shot traffic sign detection, we introduced several key modifications to enhance the model’s ability to generalize from a small number of examples. First, we keep the entire network unfrozen during training, allowing all layers to adapt to the new data and enabling more effective learning of domain-specific features. Second, we augment the training data by creating a pseudo-support set, which combines original samples with their augmented versions, increasing diversity and improving the model’s resilience to real-world variations. To further mitigate the high intra-class variance inherent in few-shot scenarios, we replace the original classifier with a cosine similarity-based classifier. This replacement incorporates embedding normalization, focusing on the intrinsic relationships between instances and classes to ensure better generalization from sparse training samples.

Additionally, we fine-tune the entire network end-to-end using the MTSD dataset, ensuring that the features learned are specifically tailored to traffic signs, which present unique characteristics distinct from the general objects in COCO. The effectiveness of these strategies is evaluated by testing FUSED-Net on the target dataset across various few-shot scenarios (1, 3, 5, and 10-shot), demonstrating its suitability for the specialized task of traffic sign detection in data-scarce environments. The overview of the FUSED-Net Architecture is shown in Fig 2.

Fig 2 — The architecture begins with a Faster R-CNN model pre-trained on the MSCOCO dataset, which is then modified by incorporating a cosine similarity-based classifier for embedding normalization. This modified Faster R-CNN undergoes domain adaptation using the Merged Traffic Sign Detection Dataset (MTSD), enhancing its ability to handle diverse traffic sign appearances. In the fine-tuning stage, the model learns from a pseudo-support set generated from the target Bangladeshi Traffic Sign Detection Dataset (BDTSD). Throughout the process, all modules of the model remain unfrozen. The images from BDTSD are included with permission from the corresponding author.

Baseline faster-RCNN

To thoroughly understand the strategies we leveraged to adapt Faster R-CNN for few-shot traffic sign detection, it is crucial to first grasp the architecture of Faster R-CNN and its evolution from earlier models. The foundation of Faster R-CNN lies in the progression from RCNN (Region-based Convolutional Neural Network) [51] to Fast R-CNN [52], each of which introduced key innovations in object detection.

The object detection process traditionally involves three steps: generating region proposals, extracting feature vectors, and classifying them. Early models like RCNN used Selective Search [53] for proposal generation and CNNs for feature extraction, improving accuracy over traditional descriptors. However, RCNN relied on a pre-trained Support Vector Machine [54] for classification, which was computationally expensive due to the need for caching features from each region proposal. Fast R-CNN improved upon this by introducing the ROI pooling layer, enabling the extraction of fixed-size feature vectors from a shared feature map, allowing proposals to share computation. This innovation sped up the process and eliminated the need for caching, significantly streamlining the detection pipeline.

Building on Fast R-CNN, Faster R-CNN [23] introduced the Region Proposal Network (RPN), a fully convolutional network, designed to predict region proposals directly from the feature maps, thus integrating the proposal generation process into the neural network. This eliminated the need for external algorithms like selective search, making the detection process both faster and more accurate. Additionally, Faster R-CNN introduced a pyramid of reference anchor boxes, which allowed it to handle objects of varying scales and aspect ratios within a single framework. These multi-scale anchors enabled the sharing of features between the RPN and the classifier, further optimizing computational efficiency.

In modern implementations, Faster R-CNN typically employs deep neural networks like ResNet [55] or VGG [56] as the backbone for feature extraction, often enhanced with a Feature Pyramid Network (FPN). These backbones are crucial for generating robust, multi-scale feature representations, which are particularly important in detecting objects that vary in size and shape. The integration of the FPN allows the network to exploit features at multiple scales, improving detection accuracy for small and large objects alike.

Given the widespread success of Faster R-CNN in various object detection tasks [57–60], many FSOD frameworks [23,37,44–46], including our own, are built upon this architecture. Specifically, our approach uses a variant of Faster R-CNN provided by the Detectron2 Framework [61], with ResNet-101 and FPN as the backbone. This combination was chosen for its ability to produce deep, rich feature representations and its effectiveness in multi-scale feature extraction.

Unfrozen parameters

In deep learning, “freezing” certain layers during training is common, especially when fine-tuning pre-trained models on a new task [62]. Freezing keeps the weights of those layers fixed, updating only the unfrozen layers. This helps retain learned features while minimizing computational overhead, particularly when the new task is similar to the original task [51,63,64]. However, in cross-domain tasks like few-shot traffic sign detection, where the target domain (e.g., region-specific traffic signs) differs significantly from the base domain (e.g., general object detection), freezing layers can hinder performance [55,65,66]. The model may struggle to learn new features essential to the target domain, as frozen layers may carry biases and features from the base domain, reducing detection accuracy [67].

To overcome this limitation, we chose to keep the entire architecture unfrozen during the fine-tuning stage. This means that all components of our detector, including the Backbone, Region Proposal Network (RPN), Feature Pyramid Network (FPN), and fully connected layers, are updated during training. Given the network with layers $L_{1}, L_{2}, . . ., L_{n}$ , unfrozen training updates the weights of all the layers such that $\nabla W_{L_{i}} \neq 0$ for $i = 1, . . ., n$ . By allowing every part of the network to adapt, we ensure that the model can learn new, domain-specific features that are essential for detecting traffic signs in different environments. This approach contrasts with the conventional method of freezing certain layers and allows our model to better handle the significant differences between the base and target domains. As a result, our model demonstrates improved adaptability and accuracy in the target domain.

Pseudo support sets (PSS)

In FSOD, the limited availability of labeled training data presents a significant challenge in developing robust models. To address this, data augmentation is a common technique used to enhance the diversity and variability of the training data. In our methodology, we implemented a data augmentation strategy during the fine-tuning stage by generating a ‘Pseudo-Support Set (PSS)’ by augmenting the original support set. The augmented support set $X_{s}^{'}$ is defined in Eq (1), where $X_{s}$ represent the original support set and T represent the transformation applied:

\begin{matrix} X_{s}^{'} = X_{s} | T (X_{s}) \end{matrix}

(1)

The creation of this pseudo-support set involves applying random Color Jitter to each sample in the original support set. Here, we chose to use only Color Jitter for augmentation, guided by [68], who showed that stronger color jittering is effective in tasks requiring visual representations.

Color Jitter is a data augmentation technique that randomly alters the brightness, contrast, saturation, and hue of an image. These specific transformations were chosen because they closely mimic the kinds of variations that are commonly encountered in real-world scenarios, such as changes in lighting conditions, different times of the day, weather conditions, and even camera settings. Brightness adjustment simulates conditions ranging from very bright, sunny environments to dim, overcast scenarios. Contrast alteration affects the differentiation between light and dark areas in an image, allowing the model to recognize objects in both high-contrast situations like strong shadows, and low-contrast environments like foggy weather. Saturation changes the intensity of colors, helping the model learn to detect traffic signs that might appear more muted or vividly colored due to different camera sensors or environmental conditions. Hue adjustment shifts the overall color balance of the image, which can account for the color variances caused by different lighting sources (e.g., fluorescent vs. natural light).

The creation of PSS by supplementing original samples with their augmented versions introduces a broader range of visual characteristics into the training data. This enriched dataset helps the model recognize patterns under diverse conditions, improving generalization and reducing overfitting. Additionally, training on the PSS enhances the model’s robustness, making it less sensitive to input variations and better equipped to handle the unpredictable changes encountered in real-world traffic sign detection, leading to more reliable and accurate performance.

Embedding normalization (EN)

Embedding normalization (EN) is crucial in our method, particularly considering our few-shot learning scenario, where high intra-class variance is a challenge [37]. In few-shot scenarios, limited training examples lead to significant variations in feature representations, making the model overly sensitive to outliers and reducing generalization. To address this, EN standardizes feature representations, ensuring that all instances are on a comparable scale, which reduces intra-class variance and enhances robustness.

To incorporate EN, we integrate a cosine similarity-based classifier into FUSED-Net, which focuses on angular relationships between feature vectors, rather than magnitudes. This approach helps mitigate the impact of varying vector lengths, ensuring more accurate detection, especially in few-shot learning where limited data can lead to inconsistent feature magnitudes. EN improves generalization by reducing the influence of outliers and aligning features more effectively, allowing the classifier to distinguish between classes with minimal data. This is particularly beneficial for our application domain, few-shot traffic sign detection, where sparse and diverse data is common.

In our implementation, we normalize each embedding vector $x_{i}$ such that all instances are on a comparable scale. This standardization facilitates the effective comparison of features across different instances. The cosine similarity between the normalized embedding vector ${\hat{x}}_{i}$ and the class weight vector ${\hat{w}}_{j}$ is calculated using Eq (2).

\begin{matrix} s i m_{(x_{i}, w_{j})} = γ {\hat{x}}_{i}^{T} {\hat{w}}_{j} \end{matrix}

(2)

where,

$γ$ is a scaling factor
${\hat{x}}_{i}$ is the normalized embedding vector of $x_{i}$
${\hat{w}}_{j}$ is the normalized embedding vector of $w_{j}$

The range of cosine similarity, $s i m_{(x_{i}, w_{j})}$ varies between $- 1$ to +1. A value of +1 indicates the vectors ${\hat{x}}_{i}$ and ${\hat{w}}_{j}$ are perfectly aligned in the same direction, implying a high degree of similarity between the instance and the class. A value of $- 1$ suggests that the vectors are diametrically opposed, indicating dissimilarity. A value near 0 indicates orthogonality, meaning there is no discernible similarity between the two vectors. The value of $γ$ is determined empirically. The normalized embedding vector of $x_{i}$ and $w_{j}$ are calculated using Eq (3).

\begin{matrix} {\hat{x}}_{i} = \frac{x_{i}}{‖ x_{i} ‖}, {\hat{w}}_{j} = \frac{w_{j}}{‖ w_{j} ‖} \end{matrix}

(3)

where,

$x_{i}$ is the embedding vector of the $i^{t h}$ instance
$‖ x_{i} ‖$ is the magnitude (or length) of the vector $x_{i}$
$w_{j}$ is the weight vector of the $j^{t h}$ class
$‖ w_{j} ‖$ is the magnitude (or length) of the vector $w_{j}$

By considering cosine similarity, we ensure that the model is invariant to vector magnitudes and instead learns to prioritize the orientation of the feature vectors, which is particularly useful for few-shot learning with limited data.

Domain adaptation

In traditional machine learning, it is generally assumed that the training and testing datasets originate from the same feature space and follow similar joint probability distributions. This assumption is critical for ensuring that the model generalizes well from the training data to unseen test data. However, in many real-world scenarios, this assumption is often violated. For instance, in tasks such as traffic sign detection, the training data may consist of images collected in one country, while the test data could come from a different country with different types of signs, environmental conditions, and even visual styles. This discrepancy in data distribution can lead to a significant drop in the model’s performance when applied to new, unseen data.

To address this challenge, the concept of domain adaptation has been introduced in the field of machine learning [69]. Domain adaptation is a subset of transfer learning that specifically aims to reduce the distributional differences between a source domain (training data) and a target domain (test data). The primary objective is to enable the model to generalize effectively to the target domain, despite the differences in feature distributions.

In our approach, we employ domain adaptation by training our model on traffic signs from different countries, thereby exposing it to a diverse set of visual features and styles. This diverse training data helps minimize the distributional discrepancy between the base domain (e.g., traffic signs in the training dataset) and the target domain (e.g., traffic signs from another country with limited data). Let $P_{S} (X, Y)$ and $P_{T} (X, Y)$ represent the source and the target distributions, respectively. By training on diverse traffic sign datasets from multiple countries, our model minimizes the discrepancy $𝒟 (P_{S}, P_{T})$ . By leveraging this approach, our model becomes more generalized and robust, capable of accurately detecting and recognizing traffic signs across different domains. The effectiveness of this strategy is reflected in the improved performance of our model, which we will discuss in the following section.

Results and discussion

In this section, we present a comprehensive analysis of our proposed FUSED-Net framework. We begin with the experimental setup, which introduces the datasets used, namely Merged Traffic Sign Detection Dataset, Bangladeshi Traffic Sign Detection Dataset, and Cross-Domain Few-Shot Object Detection Benchmark, showcasing the diverse and challenging conditions under which our model was trained and tested. Additionally, we provide the implementation details, including the hyperparameters and experimental environment within the Detectron2 framework. Next, we compare FUSED-Net with the state-of-the-art models in various few-shot traffic sign detection scenarios. This discussion also includes a per-category performance analysis, examining detection accuracy across the traffic sign categories in the query set, and the statistical significance of the reported performance based on multiple independent trials. Following this, we provide a deeper understanding of FUSED-Net’s functionality through an ablation study that highlights the incremental benefits of its components and a qualitative analysis illustrating its performance in challenging scenarios using representative examples. Finally, we assess FUSED-Net’s adaptability in generalized settings showcasing its results on the Cross-Domain Few-Shot Object Detection (CD-FSOD) benchmark, which underscores the model’s robustness and versatility in diverse real-world conditions.

Experimental setup

Datasets.

Merged Traffic Sign Detection Dataset (MTSD) The key challenge of FSL models is to generate meaningful feature descriptions with minimal instances, for which it is recommended to be trained on a diverse dataset encompassing a wide range of object categories. The hypothesis is that once the model can describe objects in general, it can provide better feature descriptions even using limited samples of the target domain. However, the widely used datasets such as PASCAL VOC [70] lack traffic sign instances, and MS COCO [49] includes only a single class for “stop sign”, making them inadequate for pretraining the baseline models of few-shot traffic sign detection. Drawing inspiration from A-RPN’s FSOD dataset [44], we constructed the Merged Traffic Sign Detection Dataset (MTSD) by combining samples from three established large-scale traffic sign detection datasets: The German Traffic Sign Detection Benchmark (GTSDB) [71], The LISA Traffic Sign Dataset (LISA) [72], and DFG Traffic Sign Data Set (DFG) [25].

The GTSDB and LISA datasets feature German and American traffic signs, respectively, and together cover a total of 90 categories. Both of these datasets are widely used in the field of traffic sign detection. The DFG dataset, which is unique in its extensive collection, contains 200 categories, making it the largest among the three. In total, the MTSD dataset includes 272 categories, derived from these three sources.

Although MTSD contains 14,416 samples, the presence of multiple traffic signs within a single image leads to a total of 23,276 annotations. To prevent potential confusion in our architecture, we excluded two categories, “roundabout” and “stop,” which appeared in multiple datasets and had varying sample representations across them. This was done to avoid the possibility of conflicting data affecting model performance.

The dataset was designed with the aim of maximizing sample diversity. To this end, we selected traffic sign datasets from two different continents: Europe (GTSDB, DFG) and America (LISA). These datasets capture traffic signs developed under different regulatory standards, design philosophies, and cultural norms, thereby providing greater variation in sign appearance and representation. European traffic signs commonly employ standardized symbols, pictograms, and distinctive shapes and color schemes, whereas traffic signs in the United States more frequently incorporate textual information. By combining these geographically diverse datasets, the Merged Traffic Sign Detection Dataset (MTSD) provides a broader range of traffic sign appearances and contextual variations, which can contribute to the development of more robust traffic sign detection models.

To provide a clearer understanding of the dataset, we present the class distribution balance in MTSD in Fig 3. The dataset exhibits a highly skewed class distribution: over 85% of the categories contain fewer than 100 samples, while some categories have more than 1000 samples. This imbalance could impact model performance in few-shot learning tasks, but it does not undermine the dataset’s primary purpose. The goal of MTSD is not to offer a balanced training set but to expose the model to a broad range of traffic sign variations from different countries. This diversity is intended to promote the learning of generalized features that can be applied across various visual styles and regulatory standards.

In our proposed framework that incorporates domain adaptation, the MTSD dataset is used for the initial training of the model. This pretraining enables the system to better understand traffic sign characteristics, which enhances its ability to detect unfamiliar signs during the fine-tuning stage. Despite the dataset’s imbalanced nature, this approach helps the model generalize more effectively across domains, improving its performance on real-world traffic sign detection tasks.

Bangladeshi Traffic Sign Detection Dataset (BDTSD) The Bangladeshi Traffic Sign Detection Dataset (BDTSD) [35] introduces distinct challenges that are not found in standard benchmarks like GTSDB, primarily due to the unique geo-social characteristics of Bangladesh. This dataset comprises 2,986 samples across 15 distinct classes of Bangladeshi traffic signs, collected under a diverse array of real-world conditions. These conditions include variations in distance, occlusion, blur, geological differences, and a range of lighting scenarios, all of which are commonly encountered in practical traffic environments. While recent studies have proposed effective solutions for challenging object detection scenarios involving multi-scale objects, adverse weather conditions, occlusions, and complex background clutter [73–75], these challenges remain largely underexplored in the domain of few-shot object detection. As a result, accurately detecting traffic signs in such demanding environments continues to be a difficult task when only a limited number of annotated training samples are available. Fig 4 showcases samples that exhibit some of these challenging conditions, providing a visual representation of the complexities involved.

Fig 4 — They are characterized by red-bordered signs with numbers and pictograms. The samples highlight various detection challenges. (a) shows a traffic sign partially occluded by trees. The traffic sign in (b) depicts multiple distant traffic signs, complicating detection. (c) presents a blurry traffic sign, while (d) captures a night scene where the sign blends into the background, posing difficulty for detection. The images from BDTSD are included with permission from the corresponding author.

The BDTSD dataset was specifically chosen as the target dataset for both our ablation studies and performance comparisons with other state-of-the-art architectures. This decision was driven by its distinctiveness compared to the datasets used to compile MTSD. While the datasets in MTSD contain more standardized samples, BDTSD reflects the real-world variability and complexity found in Bangladeshi traffic sign scenarios. The aforementioned challenges such as occlusion by other vehicles or objects, signs captured at various distances and angles, and the presence of environmental conditions like rain or fog, can significantly affect the clarity of the signs. Given these unique challenges, BDTSD provides a rigorous testing ground to evaluate the generalization capabilities of our model. By using BDTSD as the target dataset, we ensure that our model is tested against a broad spectrum of conditions, allowing us to assess its robustness and adaptability to real-world situations.

Cross-Domain Few-Shot Object Detection (CD-FSOD) Benchmark In few-shot object detection, MS COCO [49] and PASCAL VOC [70] are the most commonly used benchmarks for evaluating model generalization. Authors [37,45] commonly divide PASCAL VOCs 20 categories into 15 bases and 5 novel classes across three random splits, while MS COCO is divided into 60 non-overlapping categories as base classes and 20 as novel classes. However, these benchmarks may not fully reflect real-world scenarios where base and target domain data differ significantly. In practice, few-shot learning aims to address the challenge of limited sample availability from the target domain, a situation not entirely captured by these benchmarks. Recognizing these limitations, the Cross-Domain Few-Shot Object Detection (CD-FSOD) benchmark [46] was proposed as a more realistic and challenging alternative. Unlike traditional benchmarks where both base and novel classes are drawn from the same dataset, the CD-FSOD benchmark constructs its evaluation framework by selecting base and target domain samples from entirely different datasets. This approach ensures that the base and target domains are fundamentally distinct, closely simulating real-world scenarios where the base domain may not adequately represent the target domain. Such a benchmark is critical for testing the true generalization ability of FSOD methods, especially in cases where the target domain is significantly different from the training data.

The CD-FSOD benchmark utilizes MS COCO as its base dataset, leveraging its diverse set of categories and instances. To create a more comprehensive and challenging evaluation, three target datasets from distinct domains are incorporated into the benchmark [46]. The first of these target datasets is Arthropod Taxonomy Orders Object Detection Dataset (ArTaxOr) [76], which, while composed of natural samples similar to MS COCO, is specialized in the biological domain. It contains fine-grained categories specific to arthropods, offering 7 classes with 13,991 images for training and 1,383 images for testing. This dataset is particularly valuable for evaluating a model’s ability to generalize from broad, diverse categories to fine-grained, domain-specific categories.

The second target dataset in the CD-FSOD benchmark is the Underwater Object Detection Dataset (UODD) [77], which introduces challenges from the underwater domain. The samples in UODD are characterized by poor visibility, low color contrast, and a unique visual environment that starkly contrasts with the terrestrial imagery found in MS COCO. UODD consists of 3 classes, with 3,194 images for training and 506 images for testing. This dataset tests the model’s robustness to challenging visual conditions and domain shifts that are far removed from typical object detection scenarios.

The third and final target dataset for the CD-FSOD benchmark is Dataset for Object Detection in Aerial Images (DIOR) [78], which consists of optical remote sensing images. DIOR introduces significant perspective distortion and varying scales, making it the most dissimilar to MS COCO among the three target datasets. It includes 20 classes with 18,463 training samples and 5,000 testing samples. DIOR is particularly challenging due to the overhead perspective, requiring models to adapt to entirely different viewpoints and image structures than those found in the base dataset.

By incorporating such diverse and domain-specific target datasets in our cross-domain testing, we establish a robust platform for evaluating the cross-domain generalization capability of FUSED-Net. The inclusion ensures that the framework is tested under conditions that closely mirror the complexities of real-world applications, where the training and target domains often differ significantly.

Implementation Details.

Table 1 provides an overview of the key hyperparameters used during the base training and fine-tuning stages of FUSED-Net. Unless otherwise specified, the implementation follows the standard Detectron2 Faster R-CNN training protocol. The rationale behind key implementation choices are discussed below. For further details, including the complete implementation and the datasets used, we refer readers to our publicly available source code: https://github.com/180041123-Atiq/FUSED-Net.

Table 1. Key training hyperparameters used during base pre-training and few-shot fine-tuning.

Hyperparameter	Base Pre-training	Few-shot Fine-tuning
Initial learning rate	0.0025	0.0025
Images per batch	2	2
Optimizer	SGD	SGD
Momentum	0.9	0.9
Weight Decay	$1 \times 10^{- 4}$	$1 \times 10^{- 4}$
Maximum iterations	90000	1500, 2500, 2500, 3000
Learning rate decay	60000	1000, 2000, 2000, 2500
Dataset	MTSD	BDTSD
Number of classes	272	15

Open in a new tab

Framework and Architecture Our proposed framework, FUSED-Net, leverages a variant of Faster R-CNN seamlessly integrated within the Detectron2 framework. Specifically, we employ ResNet-101 in combination with a Feature Pyramid Network (FPN) as the backbone, ensuring robust multi-scale feature extraction. Data augmentation was a critical aspect of our implementation; we applied color jitter with a probability of 0.8 (as recommended by [68]) using the transformations provided by the torchvision library [79], enhancing the diversity and resilience of the training data.

Optimization Strategy We utilized the Stochastic Gradient Descent (SGD) solver with a momentum of 0.9 and a weight decay parameter of 10^–4 to prevent overfitting. During the initial trials, we encountered NAN/INF bounding box errors due to optimization instability caused by the mismatch between the high default learning rate and the low batch size required by our single GPU environment. The small batch size of 2 introduced significant gradient noise, and the high default learning rate (0.02) exacerbated this issue, producing overly aggressive updates that led to unstable bounding box predictions and NAN/INF errors during loss computation.

To address this, we followed Detectron2 recommendations and adjusted the base learning rate and batch size to suit our computational setup. Specifically, we reduced the batch size to 2 to avoid memory errors and recalibrated the learning rate to 0.0025 using the formula provided in the Detectron2 documentation [61,80]. This adjustment stabilized the training process and ensured consistent convergence across all architectures, including our proposed one.

Scaling Factor for Classifier In our experiments, we utilized a scaling factor, $γ$ , of 20 for the cosine-similarity based classifier, following the empirically validated setting in [37]. The embedding normalization projects both the feature embeddings and classifier weights, making the classification decision dependent primarily on their angular similarity. Since the resulting cosine similarity values are bounded within $[- 1, 1]$ , the scaling factor $γ$ is applied to increase the magnitude of the logits before the softmax operation. A higher logit magnitude produces a sharper class probability distribution and stronger optimization gradients, encouraging embeddings of the same class to become more compact while increasing the angular separation between different classes. The choice of $γ = 20$ therefore provides an effective balance between improving class discriminability and reducing intra-class variance, while avoiding a substantial degradation in the detection performance of base classes. This behavior is particularly beneficial in our Bangladeshi few-shot traffic sign detection setting, where samples from the same traffic sign category can exhibit substantial variations due to illumination, viewing angle, weather conditions, motion blur, partial occlusion, sign deterioration, and background clutter. Thus, following [37], we adopt $γ = 20$ to obtain a sharper and more discriminative classification boundary while maintaining robustness under limited training examples.

Benchmarks and Metrics We benchmarked our framework against several state-of-the-art architectures, including TFA w/cos, A-RPN, DeFRCN, FRCN-ft, and CD-FSOD. To ensure fairness and consistency, we utilized 11% of the total data available in BDTSD [35] as the query set, as suggested by [37]. We employed COCO-style average precision (AP) as our evaluation metric, measured over thresholds ranging from 0.5 to 0.95 with a step size of 0.05.

Model evaluation and analysis

Performance comparison with the state-of-the-art in few-shot traffic sign detection.

As shown in Table 2, the performance of our proposed architecture was benchmarked against several state-of-the-art FSOD models on BDTSD dataset across 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

Table 2. Comparison of mAP Performance across Different Few-Shot Object Detection Models on the Bangladeshi Traffic Sign Detection Dataset (BDTSD). The experiments were done for 1-Shot, 3-Shot, 5-Shot, and 10-Shot Scenarios. Models are sorted by whether the architecture has RPN or not, and then by the average mAP ( $μ$ ) across all shots, with the best results highlighted in boldface.

Model	1-shot	3-shot	5-shot	10-shot	$μ$
One-Stage Detectors
DETR-ft [20]	2.00	6.90	8.10	13.98	7.75
YOLOv8-ft [19]	12.80	21.80	31.20	45.50	27.83
Two-Stage Detectors
TFA w/cos [37]	4.22	6.54	8.19	9.16	7.03
A-RPN [44]	8.11	9.18	10.57	12.22	10.02
DeFRCN [45]	12.53	21.93	25.84	32.28	23.15
FRCN-ft [23]	11.36	18.58	31.19	42.12	25.81
CD-FSOD [46]	12.14	19.61	31.02	43.02	26.45
FUSED-Net (Ours)	26.64	44.22	45.94	54.44	42.81

Open in a new tab

We evaluated DETR-ft and YOLOv8-ft, two advanced architectures representing single-stage detectors on the BDTSD dataset. However, these models underperformed in few-shot scenarios. DETR-ft achieved an average mAP of only 7.74, struggling particularly in the lower-shot conditions due to its heavy reliance on large datasets and prolonged training to optimize its complex attention mechanisms. Similarly, YOLOv8-ft attained an average mAP of 27.83, surpassing CD-FSOD but failing to compete with FUSED-Net. While YOLOv8-ft demonstrated reasonable performance, its lack of a dedicated region proposal network, crucial for localizing objects in data-scarce environments, limits its effectiveness in FSOD tasks. Hence, we moved on to two-stage detectors.

Despite a moderate showing, TFA w/cos achieved the lowest average mAP of 7.03. Even though its performance improved incrementally with the number of shots, it remained significantly behind other models, particularly in the higher-shot settings, underscoring its limited scalability in few-shot scenarios. This may be attributed to the frozen backbone of the architecture failing to generalize in novel settings of BDTSD.

A-RPN demonstrated better overall performance than TFA w/cos, with a notable improvement in average mAP to 10.02. While it exhibited steady gains as the shot count increased, peaking at 12.22 in the 10-shot scenario, it still fell short compared to more recent approaches. We hypothesize that the performance of A-RPN is highly contingent on the quality of the target images, which can be inconsistent in real-world scenarios like traffic sign detection. This inconsistency likely contributed to its relatively poor performance on the BDTSD, highlighting the challenges of applying attention mechanisms in the aforementioned scenario.

DeFRCN outperformed both TFA w/cos and A-RPN, with a significant leap in average mAP to 23.15. This model showed remarkable improvements in higher-shot scenarios, particularly excelling in the 5-shot and 10-shot conditions, reflecting its effectiveness in leveraging additional shot information. However, DeFRCN’s reliance on certain frozen modules, particularly in the RPN, limits its ability to improve classification scores when working with low-quality region proposals. This limitation may have hindered its ability to perform better in BDTSD, where due to the complexity of the road images, it is difficult to generate good region proposals.

With an average mAP of 25.81, FRCN-ft delivered strong performance, particularly in higher-shot scenarios. Its mAP soared from 11.36 in the 1-shot scenario to 42.12 in the 10-shot scenario, indicating robust adaptation to increased shot availability. Its simplicity and effectiveness underscore the importance of adaptable yet straightforward architectures in FSOD, especially in domains where the object instances, such as traffic signs, exhibit relatively consistent features. The significant performance gains of FRCN-ft in higher-shot scenarios suggest that a well-calibrated fine-tuning process can match, and sometimes exceed, the performance of more complex models like CD-FSOD, especially when computational efficiency is a concern.

Among the compared models, CD-FSOD achieved the second-highest performance, with an average mAP of 26.45. Its strengths were most evident in the 5-shot and 10-shot scenarios, where it attained mAPs of 31.02 and 43.02, respectively, showing its capability to scale effectively with more training data. However, as mentioned before, the complexity of the CD-FSOD approach, with its reliance on a teacher-student distillation framework, may be over-engineered for specific tasks like traffic sign detection, where objects have relatively consistent shapes and colors. This observation inspired our approach, which simplifies the FSOD process by leveraging a pseudo-support set that facilitates FUSED-Net, a model with 60 million parameters, to perform better than CD-FSOD which requires 82 million. Hence, this simplification enhances detection performance while maintaining computational efficiency.

Our proposed architecture, integrating the strengths of the state-of-the-art approaches while addressing their limitation, consistently outperformed all other models across every shot scenario, achieving the highest average mAP 42.81. Specifically, the inclusion of a cosine-similarity-based classifier, inspired by TFA w/cos, enhances embedding normalization and comparability across instances, significantly improving performance in few-shot scenarios. Additionally, we incorporate domain adaptation strategies to ensure robust generalization to the target domain. However, unlike the complex methodologies of CD-FSOD, we achieve this through a simpler, more efficient process that is particularly well-suited to the consistent nature of traffic sign detection. Our pseudo-support set technique enriches the training process by introducing necessary variability, crucial in few-shot learning, and ensures that the model is well-calibrated to the specific characteristics of BDTSD.

In conclusion, our model FUSED-Net not only achieves superior performance in terms of mAP but also exemplifies a balanced approach to FSOD, combining simplicity and robustness. Our performance gain underscores the importance of a holistic strategy that addresses both data variability through pseudo-support sets and domain specificity via adaptation techniques, positioning FUSED-Net as a leading solution in few-shot traffic sign detection.

Per-category performance analysis.

To provide a more granular evaluation of FUSED-Net, we analyze its performance across individual traffic sign categories in the BDTSD dataset. Table 3 reports the mAP values for each of the 15 categories under 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

Table 3. Comparison of mAP performance across different categories of the Bangladeshi Traffic Sign Detection Dataset (BDTSD). The experiments were conducted under 1-shot, 3-shot, 5-shot, and 10-shot scenarios. Categories are sorted alphabetically, and the best performance for each shot setting is highlighted in boldface.

Category	1-shot	3-shot	5-shot	10-shot	$μ$
Cross Roads	16.94	49.84	35.32	63.11	41.30
Narrow Bridge	32.56	38.15	53.64	77.73	50.52
No Overtaking	15.55	58.39	56.99	56.83	46.94
No Use Of Horn	28.28	35.09	55.22	41.14	39.93
Pedestrain Crossing	38.49	41.08	52.90	61.93	48.60
Road Hump	32.17	44.95	45.25	57.69	45.01
School	37.22	45.81	53.29	59.89	49.05
Sharp Bend To The Left	28.23	40.05	36.66	42.96	36.97
Sharp Bend To The Right	19.44	19.33	14.73	33.20	21.67
Side Road Left	25.44	38.21	29.50	49.78	35.73
Side Road Right	36.33	46.20	53.29	48.86	46.17
Speed Limit 40 km/h	33.23	54.35	55.49	52.50	48.89
Speed Limit 60 km/h	54.13	33.29	67.55	73.00	56.99
Speed Limit 80 km/h	27.04	54.20	56.95	64.63	50.70
U Turn	4.34	39.00	32.22	31.88	26.86
Average	30.36	42.79	46.33	54.34	43.45

Open in a new tab

The analysis reveals notable variability in performance across different categories. For instance, the model achieves high average mAP for “Speed Limit 60 km/h” (56.99) and “Narrow Bridge” (50.52), indicating the effective detection in these scenarios. However, the performance for certain categories like “U Turn” (26.68) and “Sharp Bend to the Right” (21.67) is relatively lower, suggesting that these categories present greater challenges for few-shot detection. The underperformance in the “U Turn” and “Sharp Bend to the Right” categories can be attributed to a combination of factors. These include geometric complexity and ambiguity, where subtle distinctions in directional arrows are challenging to capture with limited training data, leading to potential misclassifications. Additionally, the minimalist visual features of these signs, often lacking text or intricate patterns, make them harder to distinguish, especially under conditions like occlusion or poor lighting. Few-shot models face inherent challenges in fine-grained detection tasks, as limited samples restrict their ability to generalize nuanced differences, further compounded by category distribution imbalances. Lastly, these signs often depend on contextual cues from the road environment, which are difficult to incorporate effectively with minimal training data.

The analysis underscores that while FUSED-Net demonstrates robust performance in most categories, it also highlights the impact of category-specific characteristics on detection accuracy. This evaluation provides further insight into the strengths and potential areas of improvement for FUSED-Net, particularly in challenging traffic sign categories.

Statistical significance of the performance.

Considering the substantial performance advantage of FUSED-Net over the other SOTA architectures, we conducted seven independent trials for each of the 1-shot, 3-shot, 5-shot, and 10-shot scenarios to assess the reproducibility and reliability of our findings. We calculated the mean mAP and corresponding 95% confidence intervals (CI95) across the trials, with the results presented in Fig 5. In terms of variation, the 1-shot and 3-shot setting exhibits the larger confidence interval $\pm 1.27$ and $\pm 1.15$ respectively, indicating greater variability when fewer support examples are available, whereas the 5-shot and 10-shot setting shows the smaller variation $\pm 0.45$ and $\pm 0.75$ respectively, suggesting more stable performance across independent trials. Overall, the narrow confidence intervals relative to the corresponding mean mAP values indicate that FUSED-Net maintains stable and reproducible performance across different few-shot scenarios.

Fig 5 — Each bar represents the mean mAP for a given shot scenario, with error bars indicating the range of variability. Additionally, the quantitative variations are indicated on the right side of each bar.

To provide a more comprehensive evaluation, we also analyzed the precision, recall, and F1 scores achieved by FUSED-Net as summarized in Table 4. The precision and recall values indicate that the model maintains an effective balance between detecting relevant instances (recall) and minimizing false positives (precision) across all shot scenarios. Notably, recall exhibits slightly higher values than precision, suggesting that FUSED-Net prioritizes capturing more potential traffic sign instances, even if it occasionally predicts additional false positives. The F1-score, representing the harmonic mean of precision and recall, improves consistently as the number of shots increases, demonstrating the model’s ability to generalize better with additional data.

Table 4. Average Precision, Recall, and F1-Score of FUSED-Net on the Bangladeshi Traffic Sign Detection Dataset (BDTSD). The experiments were conducted under 1-shot, 3-shot, 5-shot, and 10-shot scenarios. Metrics are sorted alphabetically, and the best result for each metric is highlighted in boldface.

Metric	1-shot	3-shot	5-shot	10-shot
F1-Score	0.35	0.49	0.52	0.58
Precision	0.29	0.43	0.47	0.54
Recall	0.46	0.57	0.58	0.63

Open in a new tab

To further validate the significance of these findings, we performed a statistical significance analysis using paired t-tests [81] to compare FUSED-Net’s performance with the next-best-performing SOTA method (CD-FSOD). Across all scenarios, a t-statistic of 5.73 and a p-value of 0.0106 confirm that the observed performance differences are statistically significant.

The narrow confidence intervals, coupled with the precision, recall, and F1-score analysis and the statistical significance of our findings, demonstrate the high reliability of FUSED-Net’s performance. These results instill confidence in the practical applicability of our approach, particularly in real-world scenarios where data availability is often limited.

Computational cost analysis.

To further assess the practical viability of FUSED-Net, we conducted a computational cost analysis of several SOTA methods including FUSED-Net, in terms of floating-point operations (GFLOPs), average inference time (IT) per sample, and peak GPU memory (PGM). The experiments were conducted using an RTX 2000 Ada GPU with 16 GB of VRAM. The results are summarized in Table 5.

Table 5. Computational cost analysis among SOTA methods and FUSED-Net. IT and PGM denote the average inference time (in seconds) per sample and peak GPU memory (in GB), respectively.

Methods	GFLOPs	IT	PGM
TFA w/cos	276.92	0.144	0.71
A-RPN	283.89	1.059	3.26
DeFRCN	186.22	0.568	10.44
FRCN-ft	276.92	0.145	0.71
CD-FSOD	197.22	0.069	0.80
FUSED-Net	276.92	0.146	0.71

Open in a new tab

As shown in Table 5, FUSED-Net, TFA w/cos, and FRCN-ft exhibit nearly identical computational requirements across all three metrics. This is expected because these methods share the same underlying Faster R-CNN architecture, while FUSED-Net introduces only a lightweight feature fusion strategy that does not increase the overall computational complexity. Despite requiring essentially the same computational resources, FUSED-Net achieves approximately 500% and 60% higher average performance than TFA w/cos and FRCN-ft, respectively. These results indicate that the performance improvements of FUSED-Net are achieved without incurring additional computational overhead.

Among the compared methods, A-RPN exhibits the highest computational overhead, particularly in terms of inference time. Unlike conventional few-shot detectors, including FUSED-Net, which utilize the support set only during fine-tuning, A-RPN continues to use the support set during inference. Specifically, it filters query proposals based on their similarity to the support samples and subsequently employs a multi-relation matching module for final prediction. This additional inference-stage processing results in an average inference time of 1.059 seconds per sample, which is approximately 584% higher than that of FUSED-Net.

Although DeFRCN and CD-FSOD require fewer than 200 GFLOPs, their computational characteristics differ substantially. CD-FSOD benefits from its ResNet-50 backbone, resulting in relatively low computational complexity. DeFRCN, despite employing a ResNet-101 backbone, reports lower GFLOPs due to its decoupled network design. However, its Prototypical Calibration Block introduces an additional ResNet-101 network for classification refinement, resulting in approximately four times longer inference time and nearly ten times higher peak GPU memory consumption compared with FUSED-Net. In contrast, FUSED-Net maintains a peak GPU memory consumption of only 0.71 GB while achieving substantially higher performance.

CD-FSOD is the most computationally efficient method among the compared approaches, with 197.22 GFLOPs, an average inference time of 0.069 seconds per sample, and 0.80 GB of peak GPU memory. Nevertheless, without introducing additional computational overhead beyond the standard Faster R-CNN framework, FUSED-Net achieves approximately 62% higher performance than CD-FSOD. Overall, these results demonstrate that FUSED-Net provides a favorable balance between detection performance and computational efficiency, making it suitable for practical real-world deployment.

Model insights and behavior

Ablation study.

To assess the contribution of the individual components of FUSED-Net, we conducted an ablation study. The study was performed on BDTSD across 1-shot, 3-shot, 5-shot, and 10-shot scenarios. Our model integrates four key modifications: unfrozen parameters, pseudo-support set, embedding normalization, and domain adaptation. The results, presented in Table 6, demonstrate the incremental improvements each modification contributes to the overall system performance, measured by mAP.

Table 6. Ablation Study of FUSED-Net on the Bangladeshi Traffic Sign Detection Dataset (BDTSD). UP = Unfrozen Parameters, PSS = Pseudo Support Set, EN = Embedding Normalization, and DA = Domain Adaptation. The columns with $Δ$ indicate the improvement in mAP relative to the baseline model.

UP	PSS	EN	DA	1-shot	$Δ$	3-shot	$Δ$	5-shot	$Δ$	10-shot	$Δ$
✗	✗	✗	✗	10.4	–	18.8	–	26.6	–	38.6	–
✓	✗	✗	✗	11.2	0.8	20.2	1.3	28.7	2.1	40.0	1.5
✓	✓	✗	✗	11.4	1.0	21.2	2.4	29.7	3.1	42.2	3.7
✓	✓	✓	✗	13.0	2.5	22.2	3.3	30.6	4.0	43.1	4.6
✓	✓	✓	✓	26.6	16.2	44.2	25.4	45.9	19.3	54.4	15.9

Open in a new tab

Unfrozen Parameters The decision to unfreeze all layers of our architecture during fine-tuning stage played a pivotal role in improving cross-domain detection accuracy. Unlike conventional practices where certain layers are frozen to retain pre-trained knowledge, our approach allows every layer, from the Backbone to the Region Proposal Network, Feature Pyramid Network, and fully connected layers, to adjust its weights. This flexibility is helpful in scenarios where the target domain (e.g., Bangladeshi Traffic Signs) is significantly different from the base domain (e.g., general object detection datasets). As a result, the model can learn new, domain-specific features that are essential for detecting traffic signs under diverse conditions.

While the ablation study reveals that unfreezing parameters alone yields a moderate improvement of 0.8% mAP in the 1-shot scenario, this result aligns with findings from prior research [46], which demonstrated that unfrozen parameter yields varying degrees of improvement across different FSOD architectures (as illustrated in Fig 1). For example, DeFRCN, a SOTA FSOD architecture known for its generalizability, achieved only a 1.2% improvement with unfrozen parameters. Similarly, our proposed architecture, designed for high generalizability, exhibits smaller yet meaningful gains, particularly under the highly challenging 1-shot setting.

Further, the impact of unfrozen parameters becomes more apparent in the 5-shot scenario, where a 2.1% improvement in mAP is observed. This trend suggests that while the benefits of unfreezing are modest in extremely low-data regimes, they become increasingly effective as more training data becomes available. Overall, these results demonstrate that unfreezing parameters contributes meaningfully to FUSED-Net’s adaptability and performance, particularly as the difficulty of the task or the availability of data increases.

Pseudo Support Set To overcome the challenge of limited training data inherent in FSL, we introduced a pseudo-support set through data augmentation. By applying random color jitter to the original support set, we created a richer, more diverse dataset that better represents the variability encountered in real-world scenarios. This approach proved to be particularly effective, as evident from the performance gains. For example, the addition of the pseudo-support set increased the mAP by 1.0% in the 1-shot scenario and by a more pronounced 3.7% in the 10-shot scenario. The augmented support set allowed the model to generalize more effectively, improving its ability to detect traffic signs despite the limited number of examples available for training.

To validate our choice of applying only Color Jittering, we conducted additional experiments comparing various augmentation strategies, including random cropping and horizontal flipping, both individually and in combination with Color Jittering. As shown in Table 7, applying Color Jittering alone consistently achieved the highest mAP across all shot scenarios. These results validate our decision to use Color Jitter exclusively, and reinforce its effectiveness in introducing realistic variations than other techniques into the support set, making it well suited for our target domain, where traffic signs may appear under diverse real-world conditions, including varying illumination (daytime vs. nighttime) and weather conditions (strong shadows vs. foggy environments). Thus, Color Jittering provides an effective and targeted augmentation strategy for the PSS.

Table 7. Comparison of mAP Performance of FUSED-Net with different augmentation techniques on the Bangladeshi Traffic Sign Detection (BDTSD) Dataset under 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

Color Jitter	Crop	Horizontal flip	1-shot	3-shot	5-shot	10-shot	Mean
✓	✗	✗	28.22	44.49	47.23	54.95	43.72
✗	✓	✗	25.00	41.42	46.35	54.46	41.81
✗	✗	✓	23.03	36.02	42.92	54.97	39.23
✓	✓	✗	26.53	42.42	46.70	51.18	41.71
✓	✗	✓	21.53	35.75	42.81	52.87	38.29
✗	✓	✓	21.27	36.85	42.00	55.37	38.87
✓	✓	✓	23.42	40.05	44.78	50.43	39.67

Open in a new tab

Embedding Normalization Incorporating embedding normalization further refined our model’s ability to handle high intra-class variance. By standardizing the feature representations of individual instances, our approach reduces the impact of outliers and ensures that features are comparable across different instances. This normalization is enabled by a cosine similarity-based classifier, which focuses on the angular relationships between feature vectors, thereby reducing the influence of magnitude differences. The results shown in Table 6 highlight the significant improvement this modification brings, particularly in 10-shot scenarios, where mAP increased by 4.6% compared to the baseline. This demonstrates the effect of embedding normalization in enhancing the model’s robustness and generalization capabilities.

Domain Adaptation Domain adaptation had the most significant impact on our model’s performance. By training the model on traffic signs from diverse countries, we reduced the distributional variance between the source and target domains. This strategy was crucial in enabling the model to generalize effectively across different environments. The results in Table 6 underscore the effectiveness of domain adaptation, showing a dramatic increase in mAP across all shot scenarios. Even in the 1-shot scenario, the mAP jumps by 16.2% compared to the baseline. It remains substantial across higher-shot scenarios, demonstrating the effectiveness of this technique in addressing the challenges posed by domain discrepancies.

Qualitative analysis.

For the qualitative analysis, we evaluated our model’s performance by fine-tuning it on the MTSD dataset and testing it with a 5-shot scenario from BDTSD. In Figs 6 and 7, the green bounding boxes represent detections made by our model, while the blue bounding boxes denote the ground truth annotations. The labels above the blue boxes indicate the actual class labels, whereas the labels above the green boxes indicate the class labels predicted by our model. Note that, only detected boxes with confidence scores of 0.5 or higher are displayed.

As demonstrated in Fig 6(a), our model effectively detects traffic signs that are blurry and located at significant distances from the camera, a challenging scenario that often leads to detection failures in other models. Additionally, Fig 6(b) shows that our model can accurately detect and classify multiple traffic signs within a single image, even when they are positioned at varying distances and embedded in complex backgrounds. This ability to handle both single and multiple instances under challenging conditions underscores the robustness of our approach in real-world scenarios.

However, our analysis also reveals areas for improvement. As shown in Fig 7, while the model successfully detects traffic signs that are extremely distant and blurred, it sometimes fails to correctly classify them. This misclassification likely stems from the difficulty of discerning complex symbols on distant signs, a challenge even for human observers. This limitation suggests that although our model is effective in detection, the performance of the classifier could be enhanced by refining its ability to interpret subtle visual cues in challenging contexts. Future work could focus on improving the sensitivity of the classifier to these complex visual patterns, especially in low-visibility conditions, to further enhance the overall performance of the model.

Performance analysis on Cross-Domain Few-Shot Object Detection (CD-FSOD) benchmark.

To thoroughly assess the cross-domain performance of our proposed architecture, we conducted extensive experiments on the CD-FSOD benchmark, which evaluates models using base datasets from one domain (MS COCO) and novel datasets from distinct domains (ArTaxOr, UODD, DIOR). The results of these experiments, as shown in Table 8, offer insights into the efficacy of our approach compared to the state-of-the-art (SOTA) FSOD models.

Table 8. Comparison of mAP performance across various few-shot object detection models on the CD-FSOD Benchmark. Experiments were conducted under 1-shot, 5-shot, and 10-shot scenarios across the ArTaxOr, UODD, and DIOR datasets. Models are sorted by the average mAP ( $μ$ ) across all shots and datasets, with the best results highlighted in boldface.

Model	ArTaxOr			UODD			DIOR			$μ$
Model	1	5	10	1	5	10	1	5	10	$μ$
TFA w/cos [37]	0.8	3.8	7.1	2.7	6.5	7.2	3.9	10.3	13.4	6.2
A-RPN [44]	1.4	5.0	9.7	2.3	6.1	9.2	6.3	14.4	17.9	8.0
FSCE [47]	1.8	5.9	10.8	2.8	6.7	8.4	5.6	14.9	18.3	8.4
Meta-RCNN [82]	1.9	6.1	10.3	2.8	6.7	9.9	5.5	15.2	18.8	8.6
H-GCN [83]	2.0	6.3	10.9	4.5	6.2	9.4	6.0	15.4	18.7	8.8
DeFRCN [45]	2.9	8.8	14.5	3.7	8.9	10.7	8.0	17.7	20.7	10.7
FRCN-ft [23]	3.4	9.3	15.2	4.1	9.2	12.3	8.4	18.3	21.2	11.3
CD-FSOD [46]	5.1	12.5	18.1	5.9	12.2	14.5	10.5	19.1	26.5	13.8
FUSED-Net (Ours)	3.6	16.4	23.5	4.4	13.6	12.7	11.4	19.2	23.3	14.2

Open in a new tab

As discussed before, the traditional FSOD benchmarks, such as MS COCO and Pascal VOC, typically involve dividing data into base and novel sets drawn from the same domain. However, this does not reflect real-world applications where domain shifts are common. For instance, a model trained on natural images might need to detect objects in aerial or optical remote sensing images, domains with significantly different characteristics. Models fine-tuned within the same domain may perform well under such circumstances but often struggle when applied across domains. To identify whether it can excel in cross-domain settings, our proposed model, while originally intended for few-shot traffic sign detection, was subjected to rigorous testing on the CD-FSOD benchmark to evaluate its versatility.

While our model achieves the highest average mAP across all datasets and shot settings (14.2), outperforming the leading SOTA models, including CD-FSOD, it exhibits inconsistencies in the UODD dataset at the 10-shot scenario (12.7% vs CD-FSOD’s 14.5%). This discrepancy likely stems from differences between traffic sign images and underwater object images. Underwater scenes often feature clear, uniform backgrounds (e.g., blue ocean water) and slightly blurry objects, whereas traffic scenes contain highly cluttered and noisy backgrounds.

The Pseudo Support Set technique used in FUSED-Net is highly effective for challenging scenarios like traffic signs with complex surroundings, enabling it to detect such objects robustly. However, in simpler environments like underwater scenes, this method may inadvertently cause overfitting, which limits its adaptability. Despite these domain-specific challenges, our model outperforms SOTA models in overall cross-domain performance and consistently excels in 5-shot scenarios, emphasizing its utility in diverse applications.

SOTA architectures, such as TFA w/cos, A-RPN, FSCE, Metha-RCNN, H-GCN, and DeFRCN, involve freezing certain components during fine-tuning, which may stabilize the model within the same domain but often leads to poor performance in cross-domain settings. The data clearly shows that models with frozen components struggle with significant domain shifts, achieving lower mAPs across the board. For instance, TFA w/cos only achieved a 7.2 average mAP, reflecting its limited adaptability when faced with diverse domains.

Although FRCN-ft and CD-FSOD do not freeze any components, our model surpasses them in average performance. The performance disparity with FRCN-ft stems from its reliance on a limited number of labeled samples during fine-tuning, which hampers its ability to generalize in FSL. This scarcity leads to inadequate learning of novel classes, especially in complex domains such as the CD-FSOD benchmark, resulting in suboptimal performance. In contrast, our model utilizes a pseudo-support set technique to amplify the data available during fine-tuning by generating additional samples from labeled data. This broader data exposure allows our model to learn more robust features, enhancing generalization and reducing overfitting, leading to superior performance.

Our proposed FUSED-Net model even surpasses the average performance of CD-FSOD, which was designed specifically for cross-domain scenarios. This success can be attributed to FUSED-Net’s straightforward yet effective training strategy. Unlike CD-FSOD, which employs a complex two-model system with adaptive learning based on augmentations, our approach uses a single model trained with a pseudo-support set. This method not only simplifies the training process but also enhances generalization by effectively increasing the data supply during fine-tuning.

In conclusion, while FUSED-Net demonstrates slight domain-specific limitations in simpler datasets like UODD, the overall results firmly establish it as a versatile and robust architecture for cross-domain FSOD. This achievement, coupled with its simplicity and robustness, positions our approach as a highly effective solution for real-world FSOD, even in scenarios involving significant domain shifts.

Conclusion

In this work, we presented FUSED-Net, a novel Few-Shot Traffic Sign Detector, designed to address the challenges of traffic sign detection in scenarios with limited labeled data. Our model integrates domain adaptation and a pseudo-support set to significantly enhance the detection accuracy of traffic signs, where state-of-the-art methods often falter. Both approaches can be seamlessly incorporated into existing detection architectures without disrupting the training process, offering practical solutions to the data scarcity problem in traffic sign detection in remote areas. We also examined the limitations of freezing model components, a common practice in current SOTA architectures, which we found to be detrimental in few-shot detection tasks. By allowing for greater adaptability through unfrozen parameters, combined with embedding normalization, FUSED-Net demonstrates superior generalization capabilities across diverse domains. Our work not only advances the field of few-shot traffic sign detection but also highlights the importance of adaptable model components and domain-aware techniques in overcoming the challenges of few-shot learning. We hope this work inspires further innovation and exploration in few-shot learning, particularly in specialized domains like traffic sign detection, where data availability is often limited.

Data Availability

No data was generated by this study. The following existing data sources were used: 1. The German Traffic Sign Detection Benchmark from Germany, available via https://sid.erda.dk/public/archives/ff17dc924eba88d5d01a807357d6614c/published-archive.html 2. LISA Traffic Sign Dataset from the US, available via https://www.kaggle.com/datasets/mbornoe/lisa-traffic-light-dataset 3. DFG Traffic Sign Dataset from Slovenia, available via https://www.vicos.si/resources/dfg/ 4. Bangladeshi Traffic Sign Detection Dataset from Bangladesh, available via https://www.kaggle.com/datasets/atiqurrahman23/bangladeshi-traffic-sign-detection The code, processed data, and relevant resources associated with this study are also available in our GitHub repository: https://github.com/180041123-Atiq/FUSED-Net.

Funding Statement

This work was supported in part by Islamic University of Technology (IUT) Seed Grant 2022 (Grant No. REASP/IUT-RSG/2022/OL/07/011), awarded to Dr. Md. Hasanul Kabir. The funder’s website is http://iutoic-dhaka.edu/. The sponsor had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. There was no additional external funding received for this study.

References

1.Gudigar A, Chokkadi S, U R. A review on automatic detection and recognition of traffic sign. Multimed Tools Appl. 2014;75(1):333–64. doi: 10.1007/s11042-014-2293-7 [DOI] [Google Scholar]
2.Liu C, Li S, Chang F, Wang Y. Machine vision based traffic sign detection methods: review, analyses and perspectives. IEEE Access. 2019;7:86578–96. doi: 10.1109/ACCESS.2019.2924947 [DOI] [Google Scholar]
3.Eskandari Torbaghan M, Sasidharan M, Reardon L, Muchanga-Hvelplund LCW. Understanding the potential of emerging digital technologies for improving road safety. Accid Anal Prev. 2022;166:106543. doi: 10.1016/j.aap.2021.106543 [DOI] [PubMed] [Google Scholar]
4.Paul A, Chauhan R, Srivastava R, Baruah M. Advanced Driver Assistance Systems. In: International Mobility Conference. SAE Technical Paper. Society of Automotive Engineers (SAE); 2016. p. 5. Available from: https://saemobilus.sae.org/papers/advanced-driver-assistance-systems-2016-28-0223
5.Chan C-Y. Advancements, prospects, and impacts of automated driving systems. International Journal of Transportation Science and Technology. 2017;6(3):208–16. doi: 10.1016/j.ijtst.2017.07.008 [DOI] [Google Scholar]
6.Peng X, Zhou D, Sun G, Shi J, Wu L. MAD: meta adversarial defense benchmark. Sci China Inf Sci. 2026;69(6). doi: 10.1007/s11432-024-4880-x [DOI] [Google Scholar]
7.Peng X, Zhou D, Zhang J, Shi J, Sun G. RPF-MAD: A Robust Pre-Training–Fine-Tuning Algorithm for Meta-Adversarial Defense on the Traffic Sign Classification System of Autonomous Driving. Electronics. 2025;14(10). doi: 10.3390/electronics14102044 [DOI] [Google Scholar]
8.Peng X, Zhou D, Zhang J, Sun G, Ngo AT, Chattopadhyay A. NutVLM: A Self-Adaptive Defense Framework Against Full-Dimension Attacks for Vision Language Models in Autonomous Driving. In: 2026. https://doi.org/arXiv:2602.13293
9.Li Y, Wang Z, Zhang H, Yao X, Zhou Z, Cheng X. A small‐target traffic sign detection algorithm based on partial conv and atrous spatial pyramid. IET Image Processing. 2024;18(14):4639–52. doi: 10.1049/ipr2.13274 [DOI] [Google Scholar]
10.Wei H, Zhang Q, Qin Y, Li X, Qian Y. YOLOF-F: you only look one-level feature fusion for traffic sign detection. Vis Comput. 2023;40(2):747–60. doi: 10.1007/s00371-023-02813-1 [DOI] [Google Scholar]
11.Du S, Pan W, Li N, Dai S, Xu B, Liu H, et al. TSD‐YOLO: Small traffic sign detection based on improved YOLO v8. IET Image Processing. 2024;18(11):2884–98. doi: 10.1049/ipr2.13141 [DOI] [Google Scholar]
12.Qian YJ, Wang B. TSDet: A new method for traffic sign detection based on YOLOv5‐SwinT. IET Image Processing. 2023;18(4):875–85. doi: 10.1049/ipr2.12991 [DOI] [Google Scholar]
13.Kheder MQ, Mohammed AA. Retraction Note: Improved traffic sign recognition system (itsrs) for autonomous vehicle based on deep convolutional neural network. Multimed Tools Appl. 2026;85(4). doi: 10.1007/s11042-026-21526-w [DOI] [Google Scholar]
14.Kang J, Tariq S, Oh H, Woo SS. A Survey of Deep Learning-Based Object Detection Methods and Datasets for Overhead Imagery. IEEE Access. 2022;10:20118–34. doi: 10.1109/access.2022.3149052 [DOI] [Google Scholar]
15.Wang K, Zhou M, Lin Q, Niu G, Zhang X. Geometry-Guided Point Generation for 3D Object Detection. IEEE Signal Process Lett. 2025;32:136–40. doi: 10.1109/lsp.2024.3503359 [DOI] [Google Scholar]
16.Zhang X, Li X, Zhou M, Gan M, Chen CL. ASCFormer: An Adaptive Structure-Aware Cascaded Transformer for 3D Object Detection. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(2):2218–32. doi: 10.1109/TCSVT.2025.3612592 [DOI] [Google Scholar]
17.Han X, Zhang X, Niu G, Zhou M, Pan Z. Soft-label guided multi-granularity prompts learning for human-object interaction detection. Applied Soft Computing. 2026;192:114765. doi: 10.1016/j.asoc.2026.114765 [DOI] [Google Scholar]
18.Redmon J, Divvala S, Girshick R, Farhadi A. You Only Look Once: Unified, Real-Time Object Detection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 779–88. 10.1109/cvpr.2016.91 [DOI]
19.Varghese R, M S. YOLOv8: A Novel Object Detection Algorithm with Enhanced Performance and Robustness. In: 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024. 1–6. 10.1109/ADICS58448.2024.10533619 [DOI]
20.Carion N, Massa F, Synnaeve G, Usunier N, Kirillov A, Zagoruyko S. End-to-End Object Detection with Transformers. Lecture Notes in Computer Science. Springer International Publishing. 2020. p. 213–29. doi: 10.1007/978-3-030-58452-8_13 [DOI] [Google Scholar]
21.Alamgir RM, Shuvro AA, Al Mushabbir M, Raiyan MA, Rani NJ, Rahman MdM, et al. Performance Analysis of YOLO-based Architectures for Vehicle Detection from Traffic Images in Bangladesh. In: 2022 25th International Conference on Computer and Information Technology (ICCIT), 2022. 982–7. 10.1109/iccit57492.2022.10055758 [DOI]
22.Rahman R, Bin Azad Z, Bakhtiar Hasan Md. Densely-Populated Traffic Detection Using YOLOv5 and Non-maximum Suppression Ensembling. Lecture Notes on Data Engineering and Communications Technologies. Springer Singapore. 2021. p. 567–78. doi: 10.1007/978-981-16-6636-0_43 [DOI] [Google Scholar]
23.Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: Advances in Neural Information Processing Systems, 2015. 91–9. https://proceedings.neurips.cc/paper/2015/hash/14bfa6bb14875e45bba028a21ed38046-Abstract.html [DOI] [PubMed]
24.Zhu Z, Liang D, Zhang S, Huang X, Li B, Hu S. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2110–8. 10.1109/CVPR.2016.232 [DOI]
25.Tabernik D, Skocaj D. Deep Learning for Large-Scale Traffic-Sign Detection and Recognition. IEEE Trans Intell Transport Syst. 2020;21(4):1427–40. doi: 10.1109/tits.2019.2913588 [DOI] [Google Scholar]
26.Chen J, Jia K, Chen W, Lv Z, Zhang R. A real-time and high-precision method for small traffic-signs recognition. Neural Comput & Applic. 2021;34(3):2233–45. doi: 10.1007/s00521-021-06526-1 [DOI] [Google Scholar]
27.Abdel-Salam R, Mostafa R, Abdel-Gawad AH. RIECNN: real-time image enhanced CNN for traffic sign recognition. Neural Comput & Applic. 2022;34(8):6085–96. doi: 10.1007/s00521-021-06762-5 [DOI] [Google Scholar]
28.Wang J, Chen Y, Dong Z, Gao M. Improved YOLOv5 network for real-time multi-scale traffic sign detection. Neural Comput & Applic. 2022;35(10):7853–65. doi: 10.1007/s00521-022-08077-5 [DOI] [Google Scholar]
29.Gezgin H, Alkan RM. Traffic sign detection and recognition based on MMS data using YOLOv4-Tiny algorithm. Neural Comput & Applic. 2024;36(33):20633–51. doi: 10.1007/s00521-024-10279-y [DOI] [Google Scholar]
30.Zhang G, Wang Y, Li W, Fu X. Traffic sign recognition model based on scale sequence features and high-order spatial interactions. Neural Comput & Applic. 2024;37(5):3605–20. doi: 10.1007/s00521-024-10677-2 [DOI] [Google Scholar]
31.Qi G-J, Luo J. Small Data Challenges in Big Data Era: A Survey of Recent Progress on Unsupervised and Semi-Supervised Methods. IEEE Trans Pattern Anal Mach Intell. 2022;44(4):2168–87. doi: 10.1109/TPAMI.2020.3031898 [DOI] [PubMed] [Google Scholar]
32.Antonelli S, Avola D, Cinque L, Crisostomi D, Foresti GL, Galasso F, et al. Few-Shot Object Detection: A Survey. ACM Comput Surv. 2022;54(11s):1–37. doi: 10.1145/3519022 [DOI] [Google Scholar]
33.Song Y, Wang T, Cai P, Mondal SK, Sahoo JP. A comprehensive survey of few-shot learning: evolution, applications, challenges, and opportunities. ACM Computing Survey. 2023;55(13s). doi: 10.1145/3582688 [DOI] [Google Scholar]
34.Liu T, Zhang L, Wang Y, Guan J, Fu Y, Zhao J, et al. Recent Few-shot Object Detection Algorithms: A Survey with Performance Comparison. ACM Trans Intell Syst Technol. 2023;14(4):1–36. doi: 10.1145/3593588 [DOI] [Google Scholar]
35.Ashik AN, Shanto MSH, Khan RH, Kabir MH, Ahmed S. Recognizing Bangladeshi Traffic Signs in the Wild. In: 2022 25th International Conference on Computer and Information Technology (ICCIT), 2022. 1004–9. 10.1109/ICCIT57492.2022.10055612 [DOI]
36.Kohler M, Eisenbach M, Gross H-M. Few-Shot Object Detection: A Comprehensive Survey. IEEE Trans Neural Netw Learn Syst. 2024;35(9):11958–78. doi: 10.1109/TNNLS.2023.3265051 [DOI] [PubMed] [Google Scholar]
37.Wang X, Huang T, Gonzalez J, Darrell T, Yu F. In: Proceedings of Machine Learning Research, 2020. 9919–28. 10.5555/3524938.3525858 [DOI]
38.Li X, Yang X, Ma Z, Xue J-H. Deep metric learning for few-shot image classification: A Review of recent developments. Pattern Recognition. 2023;138:109381. doi: 10.1016/j.patcog.2023.109381 [DOI] [Google Scholar]
39.Zhou S, Deng C, Piao Z, Zhao B. Few-shot traffic sign recognition with clustering inductive bias and random neural network. Pattern Recognition. 2020;100:107160. doi: 10.1016/j.patcog.2019.107160 [DOI] [Google Scholar]
40.Nguyen A-KT, Tran T, Nguyen PH, Dinh VQ. Self-supervised few-shot learning for real-time traffic sign classification. Int J Adv Intell Informatics. 2024;10(1):171. doi: 10.26555/ijain.v10i1.1522 [DOI] [Google Scholar]
41.Ren X, Zhang W, Wu M, Li C, Wang X. Meta-YOLO: Meta-Learning for Few-Shot Traffic Sign Detection via Decoupling Dependencies. Applied Sciences. 2022;12(11):5543. doi: 10.3390/app12115543 [DOI] [Google Scholar]
42.Sun E, Zhou D, Tian Y, Xu Z, Wang X. Transformer-based few-shot object detection in traffic scenarios. Appl Intell. 2023;54(1):947–58. doi: 10.1007/s10489-023-05245-5 [DOI] [Google Scholar]
43.Xin Z, Chen S, Wu T, Shao Y, Ding W, You X. Few-shot object detection: Research advances and challenges. Information Fusion. 2024;107:102307. doi: 10.1016/j.inffus.2024.102307 [DOI] [Google Scholar]
44.Fan Q, Zhuo W, Tang CK, Tai YW. Few-Shot Object Detection With Attention-RPN and Multi-Relation Detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2020. p. 4012–21. Available from: https://ieeexplore.ieee.org/document/9156552 [Google Scholar]
45.Qiao L, Zhao Y, Li Z, Qiu X, Wu J, Zhang C. In: IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 8681–90. 10.1109/ICCV48922.2021.00856 [DOI]
46.Xiong W. CD-FSOD: A Benchmark For Cross-Domain Few-Shot Object Detection. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 1–5. 10.1109/icassp49357.2023.10096216 [DOI]
47.Sun B, Li B, Cai S, Yuan Y, Zhang C. FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2021. p. 7352–62. Available from: https://ieeexplore.ieee.org/document/9577907 [Google Scholar]
48.Luo C, Zhan J, Xue X, Wang L, Ren R, Yang Q. Cosine Normalization: Using Cosine Similarity Instead of Dot Product in Neural Networks. Lecture Notes in Computer Science. Springer International Publishing. 2018. p. 382–91. doi: 10.1007/978-3-030-01418-6_38 [DOI] [Google Scholar]
49.Lin T-Y, Maire M, Belongie S, Hays J, Perona P, Ramanan D, et al. Microsoft COCO: Common Objects in Context. Lecture Notes in Computer Science. Springer International Publishing. 2014. p. 740–55. doi: 10.1007/978-3-319-10602-1_48 [DOI] [Google Scholar]
50.Desai K, Parikh S, Patel K, Bide P, Ghane S. Survey of Object Detection Algorithms and Techniques. Algorithms for Intelligent Systems. Springer Singapore. 2020. p. 247–57. doi: 10.1007/978-981-15-1632-0_23 [DOI] [Google Scholar]
51.Girshick R, Donahue J, Darrell T, Malik J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 580–7. 10.1109/cvpr.2014.81 [DOI]
52.Girshick R. Fast R-CNN. In: 2015 IEEE International Conference on Computer Vision (ICCV). IEEE; 2015. p. 1440–8. Available from: https://ieeexplore.ieee.org/abstract/document/7410526/ [Google Scholar]
53.Uijlings JRR, van de Sande KEA, Gevers T, Smeulders AWM. Selective Search for Object Recognition. International Journal of Computer Vision. 2013;104(2):154–71. doi: 10.1007/s11263-013-0620-5 [DOI] [Google Scholar]
54.Cortes C, Vapnik V. Support-Vector Networks. Machine Learning. 1995;20(3):273–97. Available from: https://link.springer.com/article/10.1007/BF00994018 [Google Scholar]
55.He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 770–8. 10.1109/cvpr.2016.90 [DOI]
56.Liu S, Deng W. Very deep convolutional neural network based image classification using small training sample size. In: 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR), 2015. 730–4. 10.1109/acpr.2015.7486599 [DOI]
57.Maity M, Banerjee S, Sinha Chaudhuri S. Faster R-CNN and YOLO based vehicle detection: A survey. In: 2021. 1442–7. 10.1109/ICCMC51019.2021.9418274 [DOI]
58.Alanezi MA, Shahriar MS, Hasan MdB, Ahmed S, Sha’aban YA, Bouchekara HREH. Livestock Management With Unmanned Aerial Vehicles: A Review. IEEE Access. 2022;10:45001–28. doi: 10.1109/access.2022.3168295 [DOI] [Google Scholar]
59.Ashikur Rahman ABM, Hasan MdB, Ahmed S, Ahmed T, Ashmafee MdH, Kabir MR, et al. Two Decades of Bengali Handwritten Digit Recognition: A Survey. IEEE Access. 2022;10:92597–632. doi: 10.1109/access.2022.3202893 [DOI] [Google Scholar]
60.Bashar M, Islam S, Hussain KK, Hasan MB, Rahman ABMA, Kabir MH. In pursuit of many: A review of modern multiple object tracking systems. In: 2026. https://doi.org/arXiv:2209.04796
61.Wu Y, Kirillov A, Massa F, Lo WY, Girshick R. Detectron2. https://github.com/facebookresearch/detectron2 2019.
62.Peters ME, Ruder S, Smith NA. To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Tasks. In: Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019), 2019. 7–14. 10.18653/v1/w19-4302 [DOI]
63.Yosinski J, Clune J, Bengio Y, Lipson H. In: Advances in Neural Information Processing Systems, 2014. 3320–8. https://proceedings.neurips.cc/paper_files/paper/2014/hash/375c71349b295fbe2dcdca9206f20a06-Abstract.html
64.Razavian AS, Azizpour H, Sullivan J, Carlsson S. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014. 512–9. 10.1109/CVPRW.2014.131 [DOI]
65.Chen WY, Liu YC, Kira Z, Wang YCF, Huang JB. A Closer Look at Few-shot Classification. In: International Conference on Learning Representations; 2019. p. 1–17. Available from: https://openreview.net/forum?id=HkxLXnAcFQ
66.Long M, Cao Y, Wang J, Jordan M. Learning Transferable Features with Deep Adaptation Networks. In: Bach F, Blei D, editors. 32nd International Conference on Machine Learning. vol. 37 of Proceedings of Machine Learning Research. Lille, France: PMLR; 2015. p. 97–105. Available from: https://proceedings.mlr.press/v37/long15.html
67.Liu Z, Sun M, Zhou T, Huang G, Darrell T. Rethinking the Value of Network Pruning. In: International Conference on Learning Representations; 2019. p. 1–21. Available from: https://openreview.net/forum?id=rJlnB3C5Ym
68.Chen T, Kornblith S, Norouzi M, Hinton G. A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 37th International Conference on Machine Learning, 2020. 1597–607. https://proceedings.mlr.press/v119/chen20j.html
69.Farahani A, Voghoei S, Rasheed K, Arabnia HR. A Brief Review of Domain Adaptation. Transactions on Computational Science and Computational Intelligence. Springer International Publishing. 2021. p. 877–94. doi: 10.1007/978-3-030-71704-9_65 [DOI] [Google Scholar]
70.Everingham M, Van Gool L, Williams CKI, Winn J, Zisserman A. The Pascal Visual Object Classes (VOC) Challenge. Int J Comput Vis. 2009;88(2):303–38. doi: 10.1007/s11263-009-0275-4 [DOI] [Google Scholar]
71.Houben S, Stallkamp J, Salmen J, Schlipsing M, Igel C. Detection of traffic signs in real-world images: The German traffic sign detection benchmark. In: The 2013 International Joint Conference on Neural Networks (IJCNN), 2013. 1–8. 10.1109/ijcnn.2013.6706807 [DOI]
72.Mogelmose A, Trivedi MM, Moeslund TB. Vision-Based Traffic Sign Detection and Analysis for Intelligent Driver Assistance Systems: Perspectives and Survey. IEEE Trans Intell Transport Syst. 2012;13(4):1484–97. doi: 10.1109/tits.2012.2209421 [DOI] [Google Scholar]
73.Lei Y, Shan C, Ge W. EdgeCaseDNet: An enhanced detection architecture for edge case perception in autonomous driving. PLoS One. 2025;20(12):e0338638. doi: 10.1371/journal.pone.0338638 [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Ma T, Chen Y, Li J, Yu H, He C. SSH-YOLO: YOLOv8 improved model based on Object Detection in Complex Road Scenes. PLoS One. 2026;21(4):e0343924. doi: 10.1371/journal.pone.0343924 [DOI] [PMC free article] [PubMed] [Google Scholar]
75.Chen Y, Guo B, Bassir D, Sigama A, Zhang Y. DSE-YOLO11: Dynamic feature adaptation for key traffic element detection in complex road scenes. PLoS One. 2026;21(6):e0350328. doi: 10.1371/journal.pone.0350328 [DOI] [PMC free article] [PubMed] [Google Scholar]
76.Fatma MAM. Arthropod Taxonomy Orders Object Detection in ArTaxOr Dataset Using YOLOX. Journal of Engineering and Applied Science. 2023;70(1). doi: 10.1186/s44147-023-00284-8 [DOI] [Google Scholar]
77.Jiang L, Wang Y, Jia Q, Xu S, Liu Y, Fan X, et al. Underwater Species Detection using Channel Sharpening Attention. In: Proceedings of the 29th ACM International Conference on Multimedia, 2021. 4259–67. 10.1145/3474085.3475563 [DOI]
78.Li K, Wan G, Cheng G, Meng L, Han J. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing. 2020;159:296–307. doi: 10.1016/j.isprsjprs.2019.11.023 [DOI] [Google Scholar]
79.Marcel S, Rodriguez Y. Torchvision the machine-vision package of torch. In: 18th ACM international conference on Multimedia (MM’10). New York, NY, USA: Association for Computing Machinery; 2010. p. 1485–8. https://dl.acm.org/doi/abs/10.1145/1873951.1874254
80.76 MXLL. Floating Point Error: Predicted Boxes or Scores Contain Inf/NaN. Training Has Diverged. https://github.com/facebookresearch/detectron2/issues/1128 2020. Accessed 2026 March 20.
81.Helmert. Die Genauigkeit der Formel von Peters zur Berechnung des wahrscheinlichen Beobachtungsfehlers directer Beobachtungen gleicher Genauigkeit. Astron Nachr. 1876;88(8–9):113–31. doi: 10.1002/asna.18760880802 [DOI] [Google Scholar]
82.Han G, Huang S, Ma J, He Y, Chang S-F. Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment. AAAI. 2022;36(1):780–9. doi: 10.1609/aaai.v36i1.19959 [DOI] [Google Scholar]
83.Han G, He Y, Huang S, Ma J, Chang S-F. Query Adaptive Few-Shot Object Detection with Heterogeneous Graph Convolutional Networks. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3243–52. 10.1109/iccv48922.2021.00325 [DOI]

PLoS One. doi: 10.1371/journal.pone.0359293.r001

Decision Letter 0

Baicang Guo

23 Jul 2026

Dear Dr. Kabir,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Sep 21 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

As the corresponding author, your ORCID iD is verified in the submission system and will appear in the published article. PLOS supports the use of ORCID, and we encourage all coauthors to register for an ORCID iD and use it as well. Please encourage your coauthors to verify their ORCID iD within the submission system before final acceptance, as unverified ORCID iDs will not appear in the published article. Only the individual author can complete the verification step; PLOS staff cannot verify ORCID iDs on behalf of authors.

We look forward to receiving your revised manuscript.

Kind regards,

Baicang Guo

Academic Editor

PLOS One

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Thank you for stating in your Funding Statement: [This work was supported in part by the Islamic University of Technology (IUT) Seed Grant 2022 (Grant No. REASP/IUT-RSG/2022/OL/07/011) awarded to Dr. Md. Hasanul Kabir. The funder’s website is https://www.iutoic-dhaka.edu/. The sponsor had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript].

Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now. Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

3. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager.

4. We note that Figures 2,3,5,7 and 8 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: http://journals.plos.org/plosone/s/licenses-and-copyright.

We require you to either (1) present written permission from the copyright holder to publish these figures specifically under the CC BY 4.0 license, or (2) remove the figures from your submission:

1. You may seek permission from the original copyright holder of Figures 2,3,5,7 and 8 to publish the content specifically under the CC BY 4.0 license.

We recommend that you contact the original copyright holder with the Content Permission Form (http://journals.plos.org/plosone/s/file?id=7c09/content-permission-form.pdf) and the following text:

“I request permission for the open-access journal PLOS ONE to publish XXX under the Creative Commons Attribution License (CCAL) CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). Please be aware that this license allows unrestricted use and distribution, even commercially, by third parties. Please reply and provide explicit written permission to publish XXX under a CC BY license and complete the attached form.”

Please upload the completed Content Permission Form or other proof of granted permissions as an ""Other"" file with your submission.

In the figure caption of the copyrighted figure, please include the following text: “Reprinted from [ref] under a CC BY license, with permission from [name of publisher], original copyright [original copyright year].”

2. If you are unable to obtain permission from the original copyright holder to publish these figures under the CC BY 4.0 license or if the copyright holder’s requirements are incompatible with the CC BY 4.0 license, please either i) remove the figure or ii) supply a replacement figure that complies with the CC BY 4.0 license. Please check copyright information on all replacement figures and update the figure caption with source information. If applicable, please specify in the figure caption text when a figure is similar but not identical to the original image and is therefore for illustrative purposes only.

5. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

Additional Editor Comments:

The manuscript presents extensive quantitative evaluations, but an important weakness is the lack of sufficient qualitative visualization results. For an object detection study, representative detection examples are essential to demonstrate the practical effectiveness, robustness, and failure modes of the proposed method. At present, the manuscript relies predominantly on numerical comparisons, while visual evidence supporting the claimed performance improvements is very limited. Consequently, it is difficult to assess whether the proposed approach consistently produces better detection results in realistic scenarios or whether the reported improvements are primarily reflected in benchmark metrics.

Given that qualitative evaluation is a fundamental component of object detection research, I consider the current experimental evidence incomplete. The manuscript should provide substantially more representative detection results to support its claims. Without such evidence, the conclusions regarding the superiority and robustness of the proposed method are insufficiently supported, and I do not believe the manuscript is suitable for publication in its current form.

With respect to the references suggested during peer review, only those that are strictly relevant to the scope and technical content of the manuscript should be incorporated. Citations should be included based on scientific relevance rather than solely to accommodate reviewer requests.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #1: Yes

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #1: Yes

Reviewer #2: No

**********

3. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #1: Yes

Reviewer #2: Yes

**********

Reviewer #1: 1. The paper does not clearly articulate what is fundamentally new beyond combining these existing techniques. The term "FUSED-Net" appears to be a re-packaging of known methods rather than a conceptually novel framework.

2. The choice of only Color Jitter over other augmentations is not rigorously justified. The claim that it "closely mimics real-world variations" is subjective and unsupported by quantitative analysis.

3. The number of trials for statistical significance is mentioned but not clearly reported (e.g., standard deviations, confidence intervals in tables). The computational cost (e.g., FLOPs, inference time, GPU memory) is not reported, making it difficult to assess practical viability.

4. Some most related works should be added in this paper:

[1]Geometry-Guided Point Generation for 3D Object Detection, IEEE Signal Processing Letters, 2025. DOI: 10.1109/LSP.2024.3503359.

[2]ASCFormer: An Adaptive Strucure-aware Cascaded Transformer for 3D Object Detection. IEEE Transactions on Circuits and Systems for Video Technology, 2026. DOI: 10.1109/TCSVT.2025.3612592.

[3]Soft-Label Guided Multi-Granularity Prompts Learning for Human-Object Interaction Detection, Applied Soft Computing. DOI: 10.1016/j.asoc.2026.114765.

Reviewer #2: This paper introduces FUSED-Net, a novel approach for Few-Shot Object Detection (FSOD) specifically tailored for traffic sign detection in data-scarce environments. The authors address the impracticality of curating large-scale annotated datasets by integrating Faster R-CNN with four key components: unfrozen parameters, pseudo-support sets, embedding normalization, and domain adaptation. However, it remains many problems.

1. The paper states that Color Jitter alone yielded the highest mAP scores after comparing it with other augmentations. However, the details of these comparative experiments are not provided.

2. Include a table or list of key hyperparameters used for training (e.g., learning rate schedule, optimizer settings, number of epochs/iterations) for both pre-training and fine-tuning.

3. Ensure all figures are high-resolution and clearly convey the intended message. For Figure 8, add a more descriptive caption explaining the failure mode.

4. Regarding the "embedding normalization" using a cosine similarity-based classifier, how does the scaling factor γ (set to 20) specifically influence the separation of classes and the reduction of intra-class variance in the context of traffic signs?

5. More related works should be added in revision: https://doi.org/10.1007/s11432-024-4880-x, https://doi.org/10.3390/electronics14102044, and https://arxiv.org/abs/2602.13293v2.

**********

what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy

Reviewer #1: No

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures

You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation.

NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications.

PLoS One. 2026 Oct 7;21(10):e0359293. doi: 10.1371/journal.pone.0359293.r002

Author response to Decision Letter 1

12 Aug 2026

Thank you for the review of our manuscript and for the opportunity to submit a revision. We are very grateful for the encouraging feedback and the insightful suggestions for improvement. We have carefully considered all the comments from the editor and the reviewers and have revised the manuscript accordingly. We have structured our response below to address the provided feedback, and we believe the manuscript is now significantly stronger as a result. We would like to request the editors and reviewers to check the Response to Reviewers PDF (r1-response-july2026.pdf) for better formatting and structure. However, the responses are also provided here for convenience:

Editor Concern 1:

We prepared the manuscript using the LaTeX template (https://journals.plos.org/plosone/s/latex) provided by PLOS ONE. Following the editor's comments, we carefully reviewed the manuscript once again to ensure compliance with the journal's formatting and style requirements. To the best of our knowledge, the manuscript now fully adheres to the formatting guidelines specified by PLOS ONE. In addition, we have updated the filenames of the attached image files to conform to the journal's file-naming requirements.

Editor Concern 2:

We sincerely thank the Editor for pointing out the need to provide a complete declaration of all funding and sources of support received during the study. We have carefully considered this request and have amended the Funding Statement accordingly.

We confirm that the Islamic University of Technology (IUT) Seed Grant 2022 (Grant No. REASP/IUT-RSG/2022/OL/07/011), awarded to Dr. Md. Hasanul Kabir, is the only funding or source of financial support received for this study. No other external or internal funding was received.

The amended Funding Statement is:

"This work was supported in part by the Islamic University of Technology (IUT) Seed Grant 2022 (Grant No. REASP/IUT-RSG/2022/OL/07/011), awarded to Dr. Md. Hasanul Kabir. The funder’s website is https://www.iutoic-dhaka.edu/. The sponsor had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. There was no additional external funding received for this study."

As requested, we have also included this amended Funding Statement in the revised cover letter. We appreciate the Editor's guidance in ensuring that the funding information is complete and transparent.

Editor Concern 3:

Thank you for your comment. We have updated the corresponding author's profile in Editorial Manager to include and validate their ORCID iD, as requested. We appreciate your guidance and assistance in this regard.

Editor Concern 4:

Thank you for bringing this copyright and licensing issue to our attention. We carefully reviewed the licensing conditions of all datasets used in our manuscript and contacted the authors/copyright holders of the relevant datasets to seek permission for publication under the CC BY 4.0 license.

For the Merged Traffic Sign Detection Dataset (MTSD), the samples were drawn from the German Traffic Sign Detection Benchmark (GTSDB), the LISA Traffic Sign Dataset, and the DFG Traffic Sign Dataset. We found that these datasets have licensing conditions that are not directly compatible with the CC BY 4.0 licensing requirements of PLOS ONE: GTSDB is distributed under the CC BY-NC 4.0 license, the LISA Traffic Sign Dataset is released under an academic license agreement, and the DFG Traffic Sign Dataset is distributed under the CC BY-NC-SA 4.0 license, with commercial use requiring additional permission.

We contacted the authors/copyright holders of these datasets to request permission to publish the relevant images specifically under the CC BY 4.0 license. However, we have not received the necessary permissions at this time. Therefore, in accordance with the editorial instructions, we have modified Figure 2 to remove the copyrighted images and removed the figure containing samples from these datasets (previously Figure 3) from the manuscrip. We have also revised the corresponding discussion in the manuscript so that it no longer relies on or refers to the removed figure.

Regarding the Bangladeshi Traffic Sign Dataset (BDTSD), we contacted the corresponding author and received permission to use the dataset/images in our work. We will retain the figures utilizing BDTSD images and have added an appropriate permission statement to the captions of these figures. A copy of the written permission from the corresponding author will also be provided as an ``Other'' file with the revised submission.

We appreciate the editor's guidance, which has helped us ensure that the revised manuscript complies with PLOS ONE's copyright and licensing requirements.

Editor Concern 5:

We sincerely thank the Editor for this clarification. We carefully reviewed and evaluated all publications recommended by the reviewers and assessed their relevance to the scope and contributions of our manuscript. Based on this evaluation, we have incorporated the recommended works that provide relevant context to our discussion of recent advances in object detection and the robustness and reliability of traffic sign perception systems. The corresponding citations have been added at appropriate locations in the Introduction section of the revised manuscript.

Reviewer 1, Concern 1:

We thank the reviewer for this valuable observation. We agree that FUSED-Net is not proposed as an entirely new object detection architecture. Rather, our contribution lies in the design of a unified framework that systematically addresses the unique challenges of few-shot traffic sign detection by integrating several complementary techniques into the standard Faster R-CNN framework. While some of the individual components have been explored in other domains, their adaptation, integration, and validation for cross-domain few-shot traffic sign detection constitute the primary novelty of our work.

Specifically, the novelty of FUSED-Net can be summarized as follows:

- Fully unfrozen fine-tuning: Unlike conventional few-shot object detection approaches that freeze a portion of the network during adaptation, FUSED-Net fine-tunes the entire detector, including the backbone, FPN, RPN, and ROI heads. This enables the model to learn domain-specific representations that are crucial for adapting from general object detection datasets to region-specific traffic sign datasets.

- Pseudo Support Set (PSS): We introduce a Pseudo Support Set generated through targeted color-based augmentation of the limited support samples. To the best of our knowledge, this is the first work to explicitly employ a pseudo support set augmentation strategy for few-shot object detection. The proposed PSS increases the diversity of the support set and improves robustness against illumination, weather, and appearance variations that are frequently encountered in real-world traffic sign images.

- Embedding Normalization (EN): We incorporate a cosine similarity-based classifier with embedding normalization into the traffic sign detection framework. Although cosine similarity classifiers have been investigated in prior few-shot object detection methods, their application to cross-domain few-shot traffic sign detection has not been explored. We demonstrate that embedding normalization effectively reduces intra-class variation and improves feature discrimination among visually similar traffic sign categories.

- Domain adaptation: FUSED-Net is explicitly designed for cross-domain traffic sign detection by leveraging traffic sign datasets collected from multiple countries. This reduces the distribution gap between the source and target domains, resulting in a detector that generalizes more effectively when only a few labeled target-domain samples are available.

Therefore, the contribution of FUSED-Net extends beyond combining existing techniques. Instead, it presents a carefully designed framework in which these complementary components are adapted and integrated to address the specific challenges of few-shot traffic sign detection. Furthermore, the ablation studies presented in the manuscript demonstrate that each proposed component contributes independently to the overall performance, while their combination yields state-of-the-art detection accuracy (62\% improvement over second best method) on the Bangladeshi traffic sign detection benchmark.

Reviewer 1, Concern 2:

We thank the reviewer for raising this important concern. Initially, our intuition was that employing multiple augmentation techniques on the limited samples available in few-shot scenarios could lead to over-complication, potentially impairing the model's ability to learn. Guided by this intuition and prior literature [1], we focused on using Color Jitter, which has been shown to enhance model performance in tasks requiring visual representations. Upon empirical analysis, [1] specifically highlighted that strong Color Jitter achieves superior results, and we adopted their recommended probability of $0.8$ in our implementation.

To further address the comment, we conducted a set of experiments comparing the effectiveness of different popular augmentation strategies, including random cropping, horizontal flipping, and combinations thereof, alongside Color Jitter. The results shown in Table 1 demonstrate that Color Jitter alone outperforms other augmentation strategies and their combinations, achieving the highest average mAP across all shot scenarios. These results validate our decision to use Color Jitter exclusively, and reinforce its effectiveness in introducing realistic variations than other techniques into the support set, making it well suited for our target domain, where traffic signs may appear under diverse real-world conditions, including varying illumination (daytime vs. nighttime) and weather conditions (strong shadows vs. foggy environments). Thus, Color Jittering provides an effective and targeted augmentation strategy for the PSS.

Table 1: Comparison of mAP Performance of FUSED-Net with different augmentation techniques on the Bangladeshi Traffic Sign Detection (BDTSD) Dataset under 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

-------------|------|-----------------|--------|--------|--------|---------|------

✓ | ✗ | ✗ | 28.22 | 44.49 | 47.23 | 54.95 | 43.72

✗ | ✓ | ✗ | 25.00 | 41.42 | 46.35 | 54.46 | 41.81

✗ | ✗ | ✓ | 23.03 | 36.02 | 42.92 | 54.97 | 39.23

✓ | ✓ | ✗ | 26.53 | 42.42 | 46.70 | 51.18 | 41.71

✓ | ✗ | ✓ | 21.53 | 35.75 | 42.81 | 52.87 | 38.29

✗ | ✓ | ✓ | 21.27 | 36.85 | 42.00 | 55.37 | 38.87

✓ | ✓ | ✓ | 23.42 | 40.05 | 44.78 | 50.43 | 39.67

To address the reviewer's concern, we have included Table 1 in the Ablation Study, along with the corresponding analysis on Page 18 (Line 671 onwards). We have also clarified how Color Jittering is applied in the Implementation Details at Page 12 (line 463 onwards).

Reviewer 1, Concern 3:

We thank the reviewer for pointing out the need for clearer reporting of statistical significance and computational cost. We agree that these analyses strengthen the scientific rigor of the study and demonstrate the practical viability of FUSED-Net.

To address the statistical significance concern, we conducted seven independent trials of FUSED-Net and report the mean mAP with 95\% confidence intervals (CI95) for the 1-, 3-, 5-, and 10-shot settings in Table 2. The consistently narrow confidence intervals demonstrate stable performance across runs. We have also updated Figure 5 in the revised manuscript to show the quantitative variation for each bar and revised the Statistical Significance of the Performance subsection accordingly in Page 15 (Line 579 onwards).

Table 2: Mean mAP of FUSED-Net across different shot scenarios with 95% confidence intervals.

Setup | Mean | CI95

-------|-------|------

1-shot | 26.64 | ± 1.27

3-shot | 44.22 | ± 1.15

5-shot | 45.94 | ± 0.45

10-shot| 54.44 | ± 0.75

To address practical viability, we additionally report GFLOPs, average inference time (seconds/sample), and peak GPU memory (GB) for all compared methods in Table 3, measured using an RTX 2000 Ada with 16 GB VRAM. FUSED-Net, TFA w/cos, and FRCN-ft have nearly identical computational costs because they share the same Faster R-CNN architecture, while FUSED-Net introduces only a lightweight feature fusion module. In contrast, A-RPN incurs substantially higher inference time due to support-set processing during inference, while DeFRCN requires considerably more GPU memory and inference time. CD-FSOD has the lowest computational cost; nevertheless, FUSED-Net achieves approximately 62\% higher performance without additional computational overhead over the standard Faster R-CNN framework. These results demonstrate that FUSED-Net provides a favorable balance between detection performance and computational efficiency.

Table 3: Computational cost analysis among SOTA methods and FUSED-Net. IT denotes average inference time (in seconds) for a sample, and PGM denotes peak GPU memory (in GB).

Methods | GFLOPs | IT | PGM

----------|--------|-------|-----

TFA w/cos | 276.92 | 0.144 | 0.71

A-RPN | 283.89 | 1.059 | 3.26

DeFRCN | 186.22 | 0.568 | 10.44

FRCN-ft | 276.92 | 0.145 | 0.71

CD-FSOD | 197.22 | 0.069 | 0.80

FUSED-Net | 276.92 | 0.146 | 0.71

The corresponding discussion and table have been added to the Computational Cost Analysis subsection in Page 16 (Line 605 onwards).

Reviewer 1, Concern 4:

We thank the reviewer for suggesting these recent studies. Following the recommendation, we have incorporated the three suggested works into the second paragraph of the Introduction section to provide a broader discussion of recent advances in object detection, including geometry-guided point generation, structure-aware transformer architectures, and multi-granularity prompt learning. These studies help contextualize our work within the recent development of object detection techniques, while our work specifically focuses on few-shot traffic sign detection under limited labeled data.

Reviewer 2, Concern 1:

We thank the reviewer for this comment. We would like to clarify that, in response to the related concern raised by R\#1C\#2, we have added a detailed augmentation ablation study in the revised manuscript. Specifically, we compare Color Jitter, random cropping, and horizontal flipping, both individually and in combination, across the 1-shot, 3-shot, 5-shot, and 10-shot settings on the BDTSD dataset. The corresponding results are reported in Table 1, with the detailed analysis provided in the Ablation Study in Page 18 (Line 671 onwards).

Table 1: (Repeated for Convenience) Comparison of mAP Performance of FUSED-Net with different augmentation techniques on the Bangladeshi Traffic Sign Detection (BDTSD) Dataset under 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

-------------|------|-----------------|--------|--------|--------|---------|------

✓ | ✗ | ✗ | 28.22 | 44.49 | 47.23 | 54.95 | 43.72

✗ | ✓ | ✗ | 25.00 | 41.42 | 46.35 | 54.46 | 41.81

✗ | ✗ | ✓ | 23.03 | 36.02 | 42.92 | 54.97 | 39.23

✓ | ✓ | ✗ | 26.53 | 42.42 | 46.70 | 51.18 | 41.71

✓ | ✗ | ✓ | 21.53 | 35.75 | 42.81 | 52.87 | 38.29

✗ | ✓ | ✓ | 21.27 | 36.85 | 42.00 | 55.37 | 38.87

✓ | ✓ | ✓ | 23.42 | 40.05 | 44.78 | 50.43 | 39.67

As shown in the table, Color Jitter alone achieves the highest mean mAP (43.72) among all evaluated configurations, providing empirical support for our choice of augmentation strategy. We have also clarified the implementation details of Color Jitter in the Implementation Details section in Page 12 (line 463 onwards).

Thus, the requested comparative experimental details and quantitative justification have been incorporated into the revised manuscript.

Reviewer 2, Concern 2:

We thank the reviewer for emphasizing the importance of clearly reporting the training hyperparameters to improve the reproducibility of our work. To address this concern, we have added a summary of the key optimization and training settings used during both the base pre-training stage and the few-shot fine-tuning stage, as presented in Table 4. Unless otherwise specified, the implementation follows the standard Detectron2 Faster R-CNN training protocol with a ResNet-101-FPN backbone.

Table 4: Key training hyperparameters used during base pre-training and few-shot fine-tuning.

| Hyperparameter | Base Pre-training | Few-shot Fine-tuning |

| --------------------- | ----------------- | ---------------------- |

| Initial learning rate | 0.0025 | 0.0025 |

| Images per batch | 2 | 2 |

| Optimizer | SGD | SGD |

| Momentum | 0.9 | 0.9 |

| Weight Decay | 1 × 10−⁴ | 1 × 10−⁴ |

| Maximum iterations | 90000 | 1500, 2500, 2500, 3000 |

| Learning rate decay | 60000 | 1000, 2000, 2000, 2500 |

| Dataset | MTSD | BDTSD |

| Number of classes | 272 | 15 |

Both base training and fine-tuning are performed on a single GPU with limited memory, thus the batch size is reduced to two images, and the learning rate is scaled linearly to 0.0025 following the standard linear learning-rate scaling rule (https://github.com/facebookresearch/detectron2/issues/1128). The model is trained for 90,000 iterations, with the learning rate decayed at 60,000 iterations for base training. During few-shot fine-tuning, the backbone is initialized from the pre-trained Faster R-CNN ResNet-101-FPN model. The fine-tuning stage is performed for 1,500 iterations in the 1-shot setting, with the number of training iterations increasing proportionally for higher-shot settings. The learning rate is decayed at an appropriate stage during training for each setting. The corresponding training schedules and learning rate decay points for all shot configurations are reported accordingly.

Furthermore, we have also incorporated the analysis and Table 4 into the Implementation Details of the manuscript in Page 12 (Line 455 onwards), to address this concern.

Reviewer 2, Concern 3:

We sincerely thank the reviewer for pointing this out. We have retained the original resolution of the dataset images used in the manuscript without additional downsampling or quality degradation, while the majority of our figures are provided in PDF format to preserve high resolution. We have also revised the caption of Figure 7 (Figure 8 of the original manuscript) to provide a more detailed explanation of the observed failure mode, clarifying that the model successfully localizes the extremely distant traffic sign but incorrectly classifies it due to its small visual footprint and complex symbol. Moreover, the caption of Figure 5 (Figure 6 of the original manuscript) has been updated to clearly indicate that the ± values shown on the right side of each bar represent the quantitative variations corresponding to the 95\% confidence intervals obtained from the statistical significance analysis.

Reviewer 2, Concern 4:

We thank the reviewer for this insightful question regarding the role of the scaling factor (γ) in the cosine similarity-based classifier.

The scaling factor (γ) is adopted following the Frustratingly Simple Few-Shot Object Detection (TFA w/cos) framework [2], which employs a fixed scaling factor of 20. In their empirical study, the authors demonstrate that combining instance-level feature normalization with a cosine similarity-based classifier effectively reduces intra-class variance and improves the detection performance of novel classes, particularly in low-data few-shot scenarios, while maintaining the performance of base classes. Since FUSED-Net follows the same cosine similarity-based classification framework, we adopt the same empirically validated scaling factor to ensure a fair comparison with the baseline and to leverage its demonstrated effectiveness.

From a methodological perspective, embedding normalization projects both feature embeddings and classifier weights onto the unit hypersphere, causing classification to depend solely on the angular similarity between them. Consequently, the cosine similarity values are bounded within the range ([-1,1]), resulting in logits with relatively small magnitudes. The scaling factor (γ) rescales these logits before the softmax operation, producing a sharper probability distribution and stronger optimization gradients. This encourages feature embeddings from the same class to form more compact clusters while increasing the angular separation between different classes, thereby reducing intra-class variance and improving class discriminability.

This behavior is particularly beneficial for Bangladeshi few-shot traffic sign detection. Images of the same traffic sign often exhibit substantial appearance variations due to illumination changes, viewing angles, weather conditions, motion blur, partial occlusion, sign deterioration, and background clutter. At the same time, only a handful of training examples are available for each novel category. Under these conditions, embedding normalization together with the scaling factor of 20 encourages feature embeddings belonging to the same traffic sign category to remain tightly clustered while maintaining greater angular separation from visually similar categories. As a result, the classifier becomes more robust to the large intra-class variations commonly observed in Bangladeshi traffic signs and generalizes more effectively in the few-shot setting.

To address this concern, we have expanded the Implementation Details section in Page 13 of the updated manuscript (Line 478 onwards) to provide a detailed discussion of the rationale for selecting γ=20, its effect on feature normalization and class separation, and its relevance to the intra-class variations observed in Bangladeshi traffic signs.

Reviewer 2, Concern 5:

We thank the reviewer for suggesting these studies. The suggested references and corresponding discussion have been added to the first paragraph of the Introduction, following the discussion of the challenges associated with reliable traffic sign perception in real-world autonomous driving environments. Specifically, we have expanded the discussion of the challenges associated with reliable traffic-sign perception in autonomous driving by highlighting recent research on meta-adversarial defense, robust pre-training and fine-tuning for traffic sign classification, and self-adaptive defense against attacks on vision-language models. This addition provides a broader context regarding the robustness and reliability requirements of perception systems in safety-critical autonomous driving applications, while our work focuses specifically on addressing the complementary challenge of few-shot traffic sign detection under limited labeled data.

We believe these latest revisions have resulted in a much stronger and more complete manuscript. We thank you again for your time and consideration.

References

[1] Chen T, Kornblith S, Norouzi M, Hinton G. A Simple Framework for Contrastive Learning of Visual Representations. In: III HD, Singh A, editors. Proceedings of the 37th International Conference on

Machine Learning. vol. 119 of Proceedings of Machine Learning Research. PMLR; 2020. p. 1597-607. Available from: https://proceedings.mlr.press/v119/chen20j.html.

[2] Wang X, Huang T, Gonzalez J, Darrell T, Yu F. Frustratingly Simple Few-Shot Object Detection. In: III HD, Singh A, editors. 37th International Conference on Machine Learning (ICML). vol. 119

of Proceedings of Machine Learning Research. PMLR; 2020. p. 9919-28. Available from: https://proceedings.mlr.press/v119/wang20j.html. doi:10.5555/3524938.3525858.

Attachment

Submitted filename: r1-response-july2026.pdf

pone.0359293.s002.pdf^{(881.7KB, pdf)}

PLoS One. doi: 10.1371/journal.pone.0359293.r003

Decision Letter 1

Baicang Guo

13 Sep 2026

FUSED-Net: Detecting traffic signs with limited data

PONE-D-26-29998R1

Dear Dr. Kabir,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Baicang Guo

Academic Editor

PLOS One

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions??>

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #1: Yes

Reviewer #2: Yes

**********

Reviewer #1: The revised manuscript has addressed all comments, and the manuscript is now suitable for acceptance.

Reviewer #2: The authors make a good revison. All my concerns have been appropriately solved and I think it can be accepted now.

**********

what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy

Reviewer #1: No

Reviewer #2: No

**********

PLoS One. doi: 10.1371/journal.pone.0359293.r004

Acceptance letter

Baicang Guo

PONE-D-26-29998R1

PLOS One

Dear Dr. Kabir,

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team.

At this stage, our production department will prepare your paper for publication. This includes ensuring the following:

* All references, tables, and figures are properly cited

* All relevant supporting information is included in the manuscript submission,

* There are no issues that prevent the paper from being properly typeset

You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps.

Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing.

If we can help with anything else, please email us at customercare@plos.org.

Thank you for submitting your work to PLOS One and supporting open access.

Kind regards,

PLOS One Editorial Office Staff

on behalf of

Professor Baicang Guo

Academic Editor

PLOS One

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Attachment

Submitted filename: r1-response-july2026.pdf

pone.0359293.s002.pdf^{(881.7KB, pdf)}

Data Availability Statement

[pone.0359293.ref001] 1.Gudigar A, Chokkadi S, U R. A review on automatic detection and recognition of traffic sign. Multimed Tools Appl. 2014;75(1):333–64. doi: 10.1007/s11042-014-2293-7 [DOI] [Google Scholar]

[pone.0359293.ref002] 2.Liu C, Li S, Chang F, Wang Y. Machine vision based traffic sign detection methods: review, analyses and perspectives. IEEE Access. 2019;7:86578–96. doi: 10.1109/ACCESS.2019.2924947 [DOI] [Google Scholar]

[pone.0359293.ref003] 3.Eskandari Torbaghan M, Sasidharan M, Reardon L, Muchanga-Hvelplund LCW. Understanding the potential of emerging digital technologies for improving road safety. Accid Anal Prev. 2022;166:106543. doi: 10.1016/j.aap.2021.106543 [DOI] [PubMed] [Google Scholar]

[pone.0359293.ref004] 4.Paul A, Chauhan R, Srivastava R, Baruah M. Advanced Driver Assistance Systems. In: International Mobility Conference. SAE Technical Paper. Society of Automotive Engineers (SAE); 2016. p. 5. Available from: https://saemobilus.sae.org/papers/advanced-driver-assistance-systems-2016-28-0223

[pone.0359293.ref005] 5.Chan C-Y. Advancements, prospects, and impacts of automated driving systems. International Journal of Transportation Science and Technology. 2017;6(3):208–16. doi: 10.1016/j.ijtst.2017.07.008 [DOI] [Google Scholar]

[pone.0359293.ref006] 6.Peng X, Zhou D, Sun G, Shi J, Wu L. MAD: meta adversarial defense benchmark. Sci China Inf Sci. 2026;69(6). doi: 10.1007/s11432-024-4880-x [DOI] [Google Scholar]

[pone.0359293.ref007] 7.Peng X, Zhou D, Zhang J, Shi J, Sun G. RPF-MAD: A Robust Pre-Training–Fine-Tuning Algorithm for Meta-Adversarial Defense on the Traffic Sign Classification System of Autonomous Driving. Electronics. 2025;14(10). doi: 10.3390/electronics14102044 [DOI] [Google Scholar]

[pone.0359293.ref008] 8.Peng X, Zhou D, Zhang J, Sun G, Ngo AT, Chattopadhyay A. NutVLM: A Self-Adaptive Defense Framework Against Full-Dimension Attacks for Vision Language Models in Autonomous Driving. In: 2026. https://doi.org/arXiv:2602.13293

[pone.0359293.ref009] 9.Li Y, Wang Z, Zhang H, Yao X, Zhou Z, Cheng X. A small‐target traffic sign detection algorithm based on partial conv and atrous spatial pyramid. IET Image Processing. 2024;18(14):4639–52. doi: 10.1049/ipr2.13274 [DOI] [Google Scholar]

[pone.0359293.ref010] 10.Wei H, Zhang Q, Qin Y, Li X, Qian Y. YOLOF-F: you only look one-level feature fusion for traffic sign detection. Vis Comput. 2023;40(2):747–60. doi: 10.1007/s00371-023-02813-1 [DOI] [Google Scholar]

[pone.0359293.ref011] 11.Du S, Pan W, Li N, Dai S, Xu B, Liu H, et al. TSD‐YOLO: Small traffic sign detection based on improved YOLO v8. IET Image Processing. 2024;18(11):2884–98. doi: 10.1049/ipr2.13141 [DOI] [Google Scholar]

[pone.0359293.ref012] 12.Qian YJ, Wang B. TSDet: A new method for traffic sign detection based on YOLOv5‐SwinT. IET Image Processing. 2023;18(4):875–85. doi: 10.1049/ipr2.12991 [DOI] [Google Scholar]

[pone.0359293.ref013] 13.Kheder MQ, Mohammed AA. Retraction Note: Improved traffic sign recognition system (itsrs) for autonomous vehicle based on deep convolutional neural network. Multimed Tools Appl. 2026;85(4). doi: 10.1007/s11042-026-21526-w [DOI] [Google Scholar]

[pone.0359293.ref014] 14.Kang J, Tariq S, Oh H, Woo SS. A Survey of Deep Learning-Based Object Detection Methods and Datasets for Overhead Imagery. IEEE Access. 2022;10:20118–34. doi: 10.1109/access.2022.3149052 [DOI] [Google Scholar]

[pone.0359293.ref015] 15.Wang K, Zhou M, Lin Q, Niu G, Zhang X. Geometry-Guided Point Generation for 3D Object Detection. IEEE Signal Process Lett. 2025;32:136–40. doi: 10.1109/lsp.2024.3503359 [DOI] [Google Scholar]

[pone.0359293.ref016] 16.Zhang X, Li X, Zhou M, Gan M, Chen CL. ASCFormer: An Adaptive Structure-Aware Cascaded Transformer for 3D Object Detection. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(2):2218–32. doi: 10.1109/TCSVT.2025.3612592 [DOI] [Google Scholar]

[pone.0359293.ref017] 17.Han X, Zhang X, Niu G, Zhou M, Pan Z. Soft-label guided multi-granularity prompts learning for human-object interaction detection. Applied Soft Computing. 2026;192:114765. doi: 10.1016/j.asoc.2026.114765 [DOI] [Google Scholar]

[pone.0359293.ref018] 18.Redmon J, Divvala S, Girshick R, Farhadi A. You Only Look Once: Unified, Real-Time Object Detection. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 779–88. 10.1109/cvpr.2016.91 [DOI]

[pone.0359293.ref019] 19.Varghese R, M S. YOLOv8: A Novel Object Detection Algorithm with Enhanced Performance and Robustness. In: 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024. 1–6. 10.1109/ADICS58448.2024.10533619 [DOI]

[pone.0359293.ref020] 20.Carion N, Massa F, Synnaeve G, Usunier N, Kirillov A, Zagoruyko S. End-to-End Object Detection with Transformers. Lecture Notes in Computer Science. Springer International Publishing. 2020. p. 213–29. doi: 10.1007/978-3-030-58452-8_13 [DOI] [Google Scholar]

[pone.0359293.ref021] 21.Alamgir RM, Shuvro AA, Al Mushabbir M, Raiyan MA, Rani NJ, Rahman MdM, et al. Performance Analysis of YOLO-based Architectures for Vehicle Detection from Traffic Images in Bangladesh. In: 2022 25th International Conference on Computer and Information Technology (ICCIT), 2022. 982–7. 10.1109/iccit57492.2022.10055758 [DOI]

[pone.0359293.ref022] 22.Rahman R, Bin Azad Z, Bakhtiar Hasan Md. Densely-Populated Traffic Detection Using YOLOv5 and Non-maximum Suppression Ensembling. Lecture Notes on Data Engineering and Communications Technologies. Springer Singapore. 2021. p. 567–78. doi: 10.1007/978-981-16-6636-0_43 [DOI] [Google Scholar]

[pone.0359293.ref023] 23.Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: Advances in Neural Information Processing Systems, 2015. 91–9. https://proceedings.neurips.cc/paper/2015/hash/14bfa6bb14875e45bba028a21ed38046-Abstract.html [DOI] [PubMed]

[pone.0359293.ref024] 24.Zhu Z, Liang D, Zhang S, Huang X, Li B, Hu S. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2110–8. 10.1109/CVPR.2016.232 [DOI]

[pone.0359293.ref025] 25.Tabernik D, Skocaj D. Deep Learning for Large-Scale Traffic-Sign Detection and Recognition. IEEE Trans Intell Transport Syst. 2020;21(4):1427–40. doi: 10.1109/tits.2019.2913588 [DOI] [Google Scholar]

[pone.0359293.ref026] 26.Chen J, Jia K, Chen W, Lv Z, Zhang R. A real-time and high-precision method for small traffic-signs recognition. Neural Comput & Applic. 2021;34(3):2233–45. doi: 10.1007/s00521-021-06526-1 [DOI] [Google Scholar]

[pone.0359293.ref027] 27.Abdel-Salam R, Mostafa R, Abdel-Gawad AH. RIECNN: real-time image enhanced CNN for traffic sign recognition. Neural Comput & Applic. 2022;34(8):6085–96. doi: 10.1007/s00521-021-06762-5 [DOI] [Google Scholar]

[pone.0359293.ref028] 28.Wang J, Chen Y, Dong Z, Gao M. Improved YOLOv5 network for real-time multi-scale traffic sign detection. Neural Comput & Applic. 2022;35(10):7853–65. doi: 10.1007/s00521-022-08077-5 [DOI] [Google Scholar]

[pone.0359293.ref029] 29.Gezgin H, Alkan RM. Traffic sign detection and recognition based on MMS data using YOLOv4-Tiny algorithm. Neural Comput & Applic. 2024;36(33):20633–51. doi: 10.1007/s00521-024-10279-y [DOI] [Google Scholar]

[pone.0359293.ref030] 30.Zhang G, Wang Y, Li W, Fu X. Traffic sign recognition model based on scale sequence features and high-order spatial interactions. Neural Comput & Applic. 2024;37(5):3605–20. doi: 10.1007/s00521-024-10677-2 [DOI] [Google Scholar]

[pone.0359293.ref031] 31.Qi G-J, Luo J. Small Data Challenges in Big Data Era: A Survey of Recent Progress on Unsupervised and Semi-Supervised Methods. IEEE Trans Pattern Anal Mach Intell. 2022;44(4):2168–87. doi: 10.1109/TPAMI.2020.3031898 [DOI] [PubMed] [Google Scholar]

[pone.0359293.ref032] 32.Antonelli S, Avola D, Cinque L, Crisostomi D, Foresti GL, Galasso F, et al. Few-Shot Object Detection: A Survey. ACM Comput Surv. 2022;54(11s):1–37. doi: 10.1145/3519022 [DOI] [Google Scholar]

[pone.0359293.ref033] 33.Song Y, Wang T, Cai P, Mondal SK, Sahoo JP. A comprehensive survey of few-shot learning: evolution, applications, challenges, and opportunities. ACM Computing Survey. 2023;55(13s). doi: 10.1145/3582688 [DOI] [Google Scholar]

[pone.0359293.ref034] 34.Liu T, Zhang L, Wang Y, Guan J, Fu Y, Zhao J, et al. Recent Few-shot Object Detection Algorithms: A Survey with Performance Comparison. ACM Trans Intell Syst Technol. 2023;14(4):1–36. doi: 10.1145/3593588 [DOI] [Google Scholar]

[pone.0359293.ref035] 35.Ashik AN, Shanto MSH, Khan RH, Kabir MH, Ahmed S. Recognizing Bangladeshi Traffic Signs in the Wild. In: 2022 25th International Conference on Computer and Information Technology (ICCIT), 2022. 1004–9. 10.1109/ICCIT57492.2022.10055612 [DOI]

[pone.0359293.ref036] 36.Kohler M, Eisenbach M, Gross H-M. Few-Shot Object Detection: A Comprehensive Survey. IEEE Trans Neural Netw Learn Syst. 2024;35(9):11958–78. doi: 10.1109/TNNLS.2023.3265051 [DOI] [PubMed] [Google Scholar]

[pone.0359293.ref037] 37.Wang X, Huang T, Gonzalez J, Darrell T, Yu F. In: Proceedings of Machine Learning Research, 2020. 9919–28. 10.5555/3524938.3525858 [DOI]

[pone.0359293.ref038] 38.Li X, Yang X, Ma Z, Xue J-H. Deep metric learning for few-shot image classification: A Review of recent developments. Pattern Recognition. 2023;138:109381. doi: 10.1016/j.patcog.2023.109381 [DOI] [Google Scholar]

[pone.0359293.ref039] 39.Zhou S, Deng C, Piao Z, Zhao B. Few-shot traffic sign recognition with clustering inductive bias and random neural network. Pattern Recognition. 2020;100:107160. doi: 10.1016/j.patcog.2019.107160 [DOI] [Google Scholar]

[pone.0359293.ref040] 40.Nguyen A-KT, Tran T, Nguyen PH, Dinh VQ. Self-supervised few-shot learning for real-time traffic sign classification. Int J Adv Intell Informatics. 2024;10(1):171. doi: 10.26555/ijain.v10i1.1522 [DOI] [Google Scholar]

[pone.0359293.ref041] 41.Ren X, Zhang W, Wu M, Li C, Wang X. Meta-YOLO: Meta-Learning for Few-Shot Traffic Sign Detection via Decoupling Dependencies. Applied Sciences. 2022;12(11):5543. doi: 10.3390/app12115543 [DOI] [Google Scholar]

[pone.0359293.ref042] 42.Sun E, Zhou D, Tian Y, Xu Z, Wang X. Transformer-based few-shot object detection in traffic scenarios. Appl Intell. 2023;54(1):947–58. doi: 10.1007/s10489-023-05245-5 [DOI] [Google Scholar]

[pone.0359293.ref043] 43.Xin Z, Chen S, Wu T, Shao Y, Ding W, You X. Few-shot object detection: Research advances and challenges. Information Fusion. 2024;107:102307. doi: 10.1016/j.inffus.2024.102307 [DOI] [Google Scholar]

[pone.0359293.ref044] 44.Fan Q, Zhuo W, Tang CK, Tai YW. Few-Shot Object Detection With Attention-RPN and Multi-Relation Detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2020. p. 4012–21. Available from: https://ieeexplore.ieee.org/document/9156552 [Google Scholar]

[pone.0359293.ref045] 45.Qiao L, Zhao Y, Li Z, Qiu X, Wu J, Zhang C. In: IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 8681–90. 10.1109/ICCV48922.2021.00856 [DOI]

[pone.0359293.ref046] 46.Xiong W. CD-FSOD: A Benchmark For Cross-Domain Few-Shot Object Detection. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023. 1–5. 10.1109/icassp49357.2023.10096216 [DOI]

[pone.0359293.ref047] 47.Sun B, Li B, Cai S, Yuan Y, Zhang C. FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2021. p. 7352–62. Available from: https://ieeexplore.ieee.org/document/9577907 [Google Scholar]

[pone.0359293.ref048] 48.Luo C, Zhan J, Xue X, Wang L, Ren R, Yang Q. Cosine Normalization: Using Cosine Similarity Instead of Dot Product in Neural Networks. Lecture Notes in Computer Science. Springer International Publishing. 2018. p. 382–91. doi: 10.1007/978-3-030-01418-6_38 [DOI] [Google Scholar]

[pone.0359293.ref049] 49.Lin T-Y, Maire M, Belongie S, Hays J, Perona P, Ramanan D, et al. Microsoft COCO: Common Objects in Context. Lecture Notes in Computer Science. Springer International Publishing. 2014. p. 740–55. doi: 10.1007/978-3-319-10602-1_48 [DOI] [Google Scholar]

[pone.0359293.ref050] 50.Desai K, Parikh S, Patel K, Bide P, Ghane S. Survey of Object Detection Algorithms and Techniques. Algorithms for Intelligent Systems. Springer Singapore. 2020. p. 247–57. doi: 10.1007/978-981-15-1632-0_23 [DOI] [Google Scholar]

[pone.0359293.ref051] 51.Girshick R, Donahue J, Darrell T, Malik J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 580–7. 10.1109/cvpr.2014.81 [DOI]

[pone.0359293.ref052] 52.Girshick R. Fast R-CNN. In: 2015 IEEE International Conference on Computer Vision (ICCV). IEEE; 2015. p. 1440–8. Available from: https://ieeexplore.ieee.org/abstract/document/7410526/ [Google Scholar]

[pone.0359293.ref053] 53.Uijlings JRR, van de Sande KEA, Gevers T, Smeulders AWM. Selective Search for Object Recognition. International Journal of Computer Vision. 2013;104(2):154–71. doi: 10.1007/s11263-013-0620-5 [DOI] [Google Scholar]

[pone.0359293.ref054] 54.Cortes C, Vapnik V. Support-Vector Networks. Machine Learning. 1995;20(3):273–97. Available from: https://link.springer.com/article/10.1007/BF00994018 [Google Scholar]

[pone.0359293.ref055] 55.He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 770–8. 10.1109/cvpr.2016.90 [DOI]

[pone.0359293.ref056] 56.Liu S, Deng W. Very deep convolutional neural network based image classification using small training sample size. In: 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR), 2015. 730–4. 10.1109/acpr.2015.7486599 [DOI]

[pone.0359293.ref057] 57.Maity M, Banerjee S, Sinha Chaudhuri S. Faster R-CNN and YOLO based vehicle detection: A survey. In: 2021. 1442–7. 10.1109/ICCMC51019.2021.9418274 [DOI]

[pone.0359293.ref058] 58.Alanezi MA, Shahriar MS, Hasan MdB, Ahmed S, Sha’aban YA, Bouchekara HREH. Livestock Management With Unmanned Aerial Vehicles: A Review. IEEE Access. 2022;10:45001–28. doi: 10.1109/access.2022.3168295 [DOI] [Google Scholar]

[pone.0359293.ref059] 59.Ashikur Rahman ABM, Hasan MdB, Ahmed S, Ahmed T, Ashmafee MdH, Kabir MR, et al. Two Decades of Bengali Handwritten Digit Recognition: A Survey. IEEE Access. 2022;10:92597–632. doi: 10.1109/access.2022.3202893 [DOI] [Google Scholar]

[pone.0359293.ref060] 60.Bashar M, Islam S, Hussain KK, Hasan MB, Rahman ABMA, Kabir MH. In pursuit of many: A review of modern multiple object tracking systems. In: 2026. https://doi.org/arXiv:2209.04796

[pone.0359293.ref061] 61.Wu Y, Kirillov A, Massa F, Lo WY, Girshick R. Detectron2. https://github.com/facebookresearch/detectron2 2019.

[pone.0359293.ref062] 62.Peters ME, Ruder S, Smith NA. To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Tasks. In: Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019), 2019. 7–14. 10.18653/v1/w19-4302 [DOI]

[pone.0359293.ref063] 63.Yosinski J, Clune J, Bengio Y, Lipson H. In: Advances in Neural Information Processing Systems, 2014. 3320–8. https://proceedings.neurips.cc/paper_files/paper/2014/hash/375c71349b295fbe2dcdca9206f20a06-Abstract.html

[pone.0359293.ref064] 64.Razavian AS, Azizpour H, Sullivan J, Carlsson S. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014. 512–9. 10.1109/CVPRW.2014.131 [DOI]

[pone.0359293.ref065] 65.Chen WY, Liu YC, Kira Z, Wang YCF, Huang JB. A Closer Look at Few-shot Classification. In: International Conference on Learning Representations; 2019. p. 1–17. Available from: https://openreview.net/forum?id=HkxLXnAcFQ

[pone.0359293.ref066] 66.Long M, Cao Y, Wang J, Jordan M. Learning Transferable Features with Deep Adaptation Networks. In: Bach F, Blei D, editors. 32nd International Conference on Machine Learning. vol. 37 of Proceedings of Machine Learning Research. Lille, France: PMLR; 2015. p. 97–105. Available from: https://proceedings.mlr.press/v37/long15.html

[pone.0359293.ref067] 67.Liu Z, Sun M, Zhou T, Huang G, Darrell T. Rethinking the Value of Network Pruning. In: International Conference on Learning Representations; 2019. p. 1–21. Available from: https://openreview.net/forum?id=rJlnB3C5Ym

[pone.0359293.ref068] 68.Chen T, Kornblith S, Norouzi M, Hinton G. A Simple Framework for Contrastive Learning of Visual Representations. In: Proceedings of the 37th International Conference on Machine Learning, 2020. 1597–607. https://proceedings.mlr.press/v119/chen20j.html

[pone.0359293.ref069] 69.Farahani A, Voghoei S, Rasheed K, Arabnia HR. A Brief Review of Domain Adaptation. Transactions on Computational Science and Computational Intelligence. Springer International Publishing. 2021. p. 877–94. doi: 10.1007/978-3-030-71704-9_65 [DOI] [Google Scholar]

[pone.0359293.ref070] 70.Everingham M, Van Gool L, Williams CKI, Winn J, Zisserman A. The Pascal Visual Object Classes (VOC) Challenge. Int J Comput Vis. 2009;88(2):303–38. doi: 10.1007/s11263-009-0275-4 [DOI] [Google Scholar]

[pone.0359293.ref071] 71.Houben S, Stallkamp J, Salmen J, Schlipsing M, Igel C. Detection of traffic signs in real-world images: The German traffic sign detection benchmark. In: The 2013 International Joint Conference on Neural Networks (IJCNN), 2013. 1–8. 10.1109/ijcnn.2013.6706807 [DOI]

[pone.0359293.ref072] 72.Mogelmose A, Trivedi MM, Moeslund TB. Vision-Based Traffic Sign Detection and Analysis for Intelligent Driver Assistance Systems: Perspectives and Survey. IEEE Trans Intell Transport Syst. 2012;13(4):1484–97. doi: 10.1109/tits.2012.2209421 [DOI] [Google Scholar]

[pone.0359293.ref073] 73.Lei Y, Shan C, Ge W. EdgeCaseDNet: An enhanced detection architecture for edge case perception in autonomous driving. PLoS One. 2025;20(12):e0338638. doi: 10.1371/journal.pone.0338638 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0359293.ref074] 74.Ma T, Chen Y, Li J, Yu H, He C. SSH-YOLO: YOLOv8 improved model based on Object Detection in Complex Road Scenes. PLoS One. 2026;21(4):e0343924. doi: 10.1371/journal.pone.0343924 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0359293.ref075] 75.Chen Y, Guo B, Bassir D, Sigama A, Zhang Y. DSE-YOLO11: Dynamic feature adaptation for key traffic element detection in complex road scenes. PLoS One. 2026;21(6):e0350328. doi: 10.1371/journal.pone.0350328 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0359293.ref076] 76.Fatma MAM. Arthropod Taxonomy Orders Object Detection in ArTaxOr Dataset Using YOLOX. Journal of Engineering and Applied Science. 2023;70(1). doi: 10.1186/s44147-023-00284-8 [DOI] [Google Scholar]

[pone.0359293.ref077] 77.Jiang L, Wang Y, Jia Q, Xu S, Liu Y, Fan X, et al. Underwater Species Detection using Channel Sharpening Attention. In: Proceedings of the 29th ACM International Conference on Multimedia, 2021. 4259–67. 10.1145/3474085.3475563 [DOI]

[pone.0359293.ref078] 78.Li K, Wan G, Cheng G, Meng L, Han J. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing. 2020;159:296–307. doi: 10.1016/j.isprsjprs.2019.11.023 [DOI] [Google Scholar]

[pone.0359293.ref079] 79.Marcel S, Rodriguez Y. Torchvision the machine-vision package of torch. In: 18th ACM international conference on Multimedia (MM’10). New York, NY, USA: Association for Computing Machinery; 2010. p. 1485–8. https://dl.acm.org/doi/abs/10.1145/1873951.1874254

[pone.0359293.ref080] 80.76 MXLL. Floating Point Error: Predicted Boxes or Scores Contain Inf/NaN. Training Has Diverged. https://github.com/facebookresearch/detectron2/issues/1128 2020. Accessed 2026 March 20.

[pone.0359293.ref081] 81.Helmert. Die Genauigkeit der Formel von Peters zur Berechnung des wahrscheinlichen Beobachtungsfehlers directer Beobachtungen gleicher Genauigkeit. Astron Nachr. 1876;88(8–9):113–31. doi: 10.1002/asna.18760880802 [DOI] [Google Scholar]

[pone.0359293.ref082] 82.Han G, Huang S, Ma J, He Y, Chang S-F. Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment. AAAI. 2022;36(1):780–9. doi: 10.1609/aaai.v36i1.19959 [DOI] [Google Scholar]

[pone.0359293.ref083] 83.Han G, He Y, Huang S, Ma J, Chang S-F. Query Adaptive Few-Shot Object Detection with Heterogeneous Graph Convolutional Networks. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3243–52. 10.1109/iccv48922.2021.00325 [DOI]

PERMALINK

FUSED-Net: Detecting traffic signs with limited data

Md Atiqur Rahman

Nahian Ibn Asad

Md Mushfiqul Haque

Md Bakhtiar Hasan

Sabbir Ahmed

Md Hasanul Kabir

Roles

Abstract

Introduction

Related works

Fig 1. Effect of Parameter Freezing on Cross-domain Few-Shot Object Detection Performance.

Proposed methodology

Fig 2. Overview of the FUSED-Net architecture.

Baseline faster-RCNN

Unfrozen parameters

Pseudo support sets (PSS)

Embedding normalization (EN)

Domain adaptation

Results and discussion

Experimental setup

Datasets.

Fig 3. The dataset statistics of MTSD.

Fig 4. Samples from the Bangladeshi Traffic Sign Detection Dataset (BDTSD).

Implementation Details.

Table 1. Key training hyperparameters used during base pre-training and few-shot fine-tuning.

Model evaluation and analysis

Performance comparison with the state-of-the-art in few-shot traffic sign detection.

Per-category performance analysis.

Statistical significance of the performance.

Fig 5. Average mAP across different shot scenarios with 95% confidence intervals.

Computational cost analysis.

Table 5. Computational cost analysis among SOTA methods and FUSED-Net. IT and PGM denote the average inference time (in seconds) per sample and peak GPU memory (in GB), respectively.

Model insights and behavior

Ablation study.

Table 6. Ablation Study of FUSED-Net on the Bangladeshi Traffic Sign Detection Dataset (BDTSD). UP = Unfrozen Parameters, PSS = Pseudo Support Set, EN = Embedding Normalization, and DA = Domain Adaptation. The columns with Δ indicate the improvement in mAP relative to the baseline model.

Table 7. Comparison of mAP Performance of FUSED-Net with different augmentation techniques on the Bangladeshi Traffic Sign Detection (BDTSD) Dataset under 1-shot, 3-shot, 5-shot, and 10-shot scenarios.

Qualitative analysis.

Fig 6. Qualitative examples of successful traffic sign detection and classification by our model FUSED-Net on the BDTSD.

Fig 7. An example of a successful traffic sign detection but incorrect classification by FUSED-Net on the BDTSD dataset.

Performance analysis on Cross-Domain Few-Shot Object Detection (CD-FSOD) benchmark.

Conclusion

Data Availability

Funding Statement

References

Decision Letter 0

Baicang Guo

Roles

Author response to Decision Letter 1

Decision Letter 1

Baicang Guo

Roles

Acceptance letter

Baicang Guo

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases

Table 6. Ablation Study of FUSED-Net on the Bangladeshi Traffic Sign Detection Dataset (BDTSD). UP = Unfrozen Parameters, PSS = Pseudo Support Set, EN = Embedding Normalization, and DA = Domain Adaptation. The columns with $Δ$ indicate the improvement in mAP relative to the baseline model.