CTX-UXO: Contextual visual dataset for localisation, classification and identification of unexploded ordnance

Abstract

Unexploded Ordnances (UXOs) are explosive weapons that failed to detonate during deployment, posing significant threats to civilian safety and post-conflict recovery. The Contextual Vision for Unexploded Ordnances (CTX-UXO) dataset addresses this challenge by providing 3520 high-resolution RGB images, median size 2124 px square, containing UXOs collected over four years across all seasons and weather conditions. The dataset encompasses 15 449 annotated instances labeled as UXOs, with twelve subcategories for identification tasks, in which mortar bombs, projectiles and grenades constitute the predominant classes, supporting systems research and development in UXO classification, detection, instance segmentation and related applications, and providing multiple labelling formats such as YOLO and COCO. For the detection task including binary classification, an end-to-end solution for training, validation and testing is available as a reference by presenting a neural network optimized for edge deployment, Faster-RCNN with MobileNet V3 backbone, deployed for real-time inference with reduced costs on Nvidia Jetson Orin, achieving up to 23 frames per second with 62.18% Ap50 for UXO localisation and binary classification.

Background & Summary

Unexploded ordnance (UXO) refers to explosive devices such as bombs, projectiles, grenades that failed to detonate when deployed and still pose a risk of explosion. Contrary to prevailing assumptions, unexploded ordnance (UXOs) may pose greater risks than newly manufactured ordnance due to their potentially compromised or active arming mechanisms. These threats may retain their hazardous nature for several decades after their initial deployment or abandonment. This issue is encountered in all countries worldwide that have participated in conflicts over the last two centuries, including both World Wars, where a high volume of unexploded ordnance were left behind. According to estimates provided by the Red Cross¹, more than 120 million munitions affect the daily lives of civilian society, and 20 000 people are harmed annually, of which 80% are civilians. In addition to old munitions, the current global geopolitical context highlights the importance of recently post-conflict recovery activities. In the manuscript by Hooper et al.² it is provided that the physical properties of the environments make the mission of detecting UXOs more difficult, a leading aspect in data acquisition, curation and preprocessing for the current study. Given the challenges in UXO detection and the associated risks, the European Union has initiated new research programs³ aimed at securing continuous support for UXO clearance activities. The objective of the EU is to establish safe environments for civilian populations and to reinstate normal socio-economic activities for the impacted communities. Such an emergency situation requires a representative dataset because emergency management itself is a complex task⁴.

The integration of representative field-collected datasets with machine learning algorithms and architectures for addressing emergency situations has been established as a successful methodology in recent years, providing a robust foundation for enhancing emergency management⁴. This integration has facilitated advancements in the methods employed for data acquisition and processing in the context of emergency situations. One such dataset is introduced in recent work by Nguyen et al.⁵, which emphasises the importance of achieving the data directly from the operators, in their case from fire department personnel. It is further underscored that data acquisition in emergency scenarios occurs over extended durations due to the inherent unpredictability and non-linear temporal distribution of such situations, which renders this task labour-intensive and time-consuming. As a reference point, the study by Nguyen et al.⁵ compiles 3 124 reports of emergency situations gathered over several years, spanning all four seasons, characterised by an unequal distribution of instance captures, dictated by the actual circumstances encountered on the ground. The CTX-UXO dataset presents a comparable scale, encompassing 3 520 images with 15 449 labelled UXOs instances collected from 2021 to 2025, across all four seasons. Like the aforementioned dataset⁵, it exhibits a higher number of instances during the warmer months, correlating with increased human activity. As context, UXOs are discovered by civilians during agricultural work or during the construction of new buildings. At the same time, civil protection teams face difficulties or are prohibited from intervening during the night. In another data descriptor manuscript⁶, the authors captured approximately 1 000 field images of leafy spurge (Euphorbia esula), a highly invasive and environmentally harmful species that directly affects agricultural activities and indirectly impacts daily life. Similar to the cited work, CTX-UXO aims to mitigate the harmful effects of an intruder, in this case war remnants, on the environment and local communities. The work by Doherty et al.⁶ highlights the importance of having a representative field dataset so that machine learning algorithms can effectively address issues that impact human life. Furthermore, it presents the significance of technical validation using contemporary architectures. A similar approach will be implemented in the present work, with technical validation performed using a modern architecture deployed to an embedded hardware platform, Nvidia Jetson Orin, as a two-stage detection task⁷. In the recent study of Qin et al.⁸, the authors addressed emergency management of recent Canadian wildfires by creating an early fire/smoke detection framework. For this work, they utilised a heterogeneous dataset composed of proprietary wildfire images and the public KITTI dataset⁹. As main characteristics, the importance of using unmanned aerial-ground systems to reduce the risks associated with operators - in this case, firefighters, stands out. Even though unmanned aerial vehicles are used, images for training, validation and inference testing are generally captured from the lateral planes of wildfires since top view capture is not favourable due to forest vegetation. Additionally, they highlight the use of You Only Look Once models¹⁰ and the importance of having real-time capabilities. Comparable methods, including investigations into the capabilities of YOLO¹⁰ models with inference performed on drones¹¹ or a client-server pipeline¹², have previously been employed by the authors of the present study for the mitigation of unexploded ordnance risks. All these approaches, which are well established within the scientific community, confirm that the methodology underpinning CTX UXO is suitable for this specific context.

The detection of Unexploded Ordnance (UXO) has garnered significant interest within research communities, leading to the implementation of various methodological approaches. Anderson Lebadd et al.¹³ have contributed to this field by developing a dataset comprised of 560 UXO instances (with 420 designated for training and 140 for validation) to facilitate model training. Their work has effectively demonstrated the efficacy of Deep Neural Networks (DNNs) in accurately classifying munitions, achieving an accuracy rate of 97.1%. The images they acquired and labelled each contain a single unexploded ordnance, placed on a uniformly coloured background, with the UXOs captured from eight different angles. The methodology related to angle variations was also applied in CTX-UXO data acquisition whenever the situation allowed these operations to be performed safely. In other scientific papers^2,14, researchers have approached the UXO detection problem using signals recorded from metal detectors and Ground Penetrating Radar (GPR) in the absence of a dedicated UXO dataset for image processing tasks. While their solutions are viable, the CTX-UXO dataset aims to fill this gap by providing a complete end-to-end identification solution. The objective of this manuscript and the suggested approach, which utilizes image processing and neural networks, is not to replace techniques based on signal interpretation, but to act as a supplementary method to bolster current capabilities. Other researchers¹⁵ have attempted to use public repositories such as Collective Awareness UXO¹⁶, but unfortunately, these repositories contain mock-ups / weapons in perfect condition on white backgrounds or UXO technical drawings (learning manuals for operators of explosive ordnance disposal), making them less robust for real-world scenarios.

Figure 1 presents the data acquisition methodology as continuous and recurrent activity over time. In creating the dataset and studying how it can be implemented in viable solutions, we have implemented two types of approaches: model-centric AI and data-centric AI. Although these can typically be addressed separately¹⁷, in the current study both methods were used in parallel, using the predictions provided by the optimised models as expected to perform in a real-world scenario. The models and metric results were used to better understand the dataset, including a preliminary analysis¹⁸ of false positive and false negative cases. Regarding data-centric AI strategies, the main approaches consisted of deduplication, filtering out low-quality (blurred) images, and the use of data understanding tools¹⁹ and algorithms²⁰. For inference data development as defined by Daochen Zha et al., granular model approaches^12,18 were implemented to take advantage of specific model capabilities in certain inputs such as combining classes in different sequences¹² and the ensemble of models¹¹ with multiple voting schemes²¹ that leverage strategies that include non-maximum suppression or its weighted variants and weighted box fusion. The labelling process was performed supervised by a trained operator. The generation of masks for instance segmentation tasks was supported by the usage of Segment Anything Model²². Periodically, depending on the results obtained by the models, Explainable AI methods such as LIME²³, as well as curation platforms such as Fiftyone Voxel¹⁹ and Image DeDup²⁰, were applied. The process is ongoing and has previously been validated through four published works^11,12,18,24. This document will thoroughly investigate each of these topics. Specifically, the Data Acquisition section will elaborate on the methodology used for capturing images. Following this, the Dataset Construction subsection will provide an in-depth explanation of the techniques employed for labeling and the procedures used for eliminating artefacts. Furthermore, to offer additional support, the Preprocessing and Image Enhancement section will elucidate the implementation of various technical approaches and demonstrate how alternative methods may be utilized to enhance the robustness of the models.

Fig. 1.

CTX-UXO Dataset Acquisition, Construction and Study Methodology.

Driven by the necessity, a new dataset, Contextual Vision for Unexploded Ordnances (CTX-UXO), was created to mitigate the risk associated with unexploded ordnances, a problem which was not addressed enough before using semi-supervised or supervised algorithms and architectures mainly because of the lack of a representative dataset. CTX-UXO is periodically analysed and updated, and new image processing methodologies are tested and disseminated in scientific articles. Up to this point, the main classes of munitions that can be encountered in the field have been identified: mortar bombs, projectiles, and grenades. As a validation that the dominant classes of the dataset truly represent the real situation on the ground, the same classes of UXOs are also explicitly presented in the recent study conducted by the Red Cross Association¹. Additionally, for further studies, the CTX-UXO dataset also contains instances of classes such as aviation bombs, rocket-propelled grenades, land and sea mines, antisubmarine bombs, landmines, rockets, common cartridges, and fuses. Of course, there can be modified or rare variants that combine design elements from multiple types of munitions, such as modified mortar bombs that can be used as aviation bombs¹² - such a scenario was captured in the CTX-UXO dataset. However, these are edge cases, and the recent works of the authors ^11,12,18,24 address this aspect. Such an edge case will be explored in the Technical Validation Section of this paper.

The motivation for creating the Contextual Vision for Unexploded Ordnances (CTX-UXO) dataset stems from several critical factors. Firstly, the lack of comprehensive and representative datasets has been a significant barrier to developing effective UXO detection systems. Although image processing techniques and deep learning have shown promising results in various domains⁴, semisupervised or supervised applications to UXO detection / identification have been limited by the absence of datasets that capture the diverse real-world scenarios where UXOs are typically found. Mostly, UXOs are found in trenches, covered by mud, in different states (oxidised, incomplete, partially buried) and in uncontrolled/various environments (rain, snow, sun, etc.). Secondly, the humanitarian impact of UXOs in post-conflict regions requires the development of reliable detection systems that can operate under challenging field conditions². Thirdly, the need for real-time, resource-efficient solutions that can be deployed on reduced cost devices requires datasets that support the development and validation of such systems, with representative and studied image preprocessing steps - which are also explored in this manuscript. Such a solution, dedicated to saving human lives, should be accessible to everyone and require low-cost development and deployment. The CTX-UXO dataset offers significant potential for various machine learning applications in UXO localisation, classification, and identification. The comprehensive coverage of different UXO types, states, and environmental conditions makes it particularly suitable for deep learning architectures, as it was already tested in the past by the authors. The high-resolution images, with a median size of 2124 px square, and detailed annotations (segmentation masks, bounding boxes, class labels, etc.) enable effective feature learning and multi-approach applications. Additionally, the coverage of various environmental conditions in the dataset supports models in generalising across different lighting, weather, and background contexts. The repositories will be periodically updated by a UXO specialist to ensure that the dataset remains relevant and up-to-date.

The potential reuse value of the dataset has been demonstrated by the authors in several studies. Initial research¹² using the YOLOv8 X architecture (largest model), including a One vs. All training and inference methodology, achieved up to 80.8% mAP for binary classification and 90.6% mAP for identification tasks (multiclass detection), establishing a strong baseline for UXO detection. From the study presented, the preprocessing techniques that can be used during training stand out, the most recommended being copy-pasting on a new background²⁵. This technique will be detailed in the Preprocessing and Image Enhancement section. Another important aspect is that a granular approach, such as dividing the dataset into multiple subcategories (One vs. All approach), can improve the results for specific types of rare UXOs.

Subsequent work¹¹, starting from RetinaNet²⁶, focused on model compression and quantisation techniques, achieving real-time performance on the Nvidia Jetson Xavier NX embedded platform with 71.7% recall for UXO detection and up to 28 frames per second, demonstrating the potential of the dataset for embedded deployment. Furthermore, it has been demonstrated that Jetson devices offer the capability to implement the UXO detection solution in real-time on unmanned devices, such as EOD ground vehicles (UGVs) and drones (UAVs). The work also includes a brief analysis of loss functions, with focal loss providing improvements in handling dense instances (overlapping boxes) and the presence of negative instances (background).

In the journal article¹⁸, the authors introduce a multistep ensemble strategy for UXO detection, utilising iterative fine-tuning on false negative samples to train expert RT-DETR R18 models. These models are combined using affirmative voting and various post-processing techniques, including Non-Maximum Suppression (NMS), Soft-NMS, and Non-Maximum Weighted (NMW). The most effective configuration (Affirmative + NMS) achieves a mean mean precision (mAP) of 52. 4% and a recall of 86. 2%, improving from the baseline by +4.5% mAP and +3.9% recall. Compared to a larger RT-DETR R101 model (76M parameters), the ensemble reaches nearly mAP (0.1%) but with +11.1% higher recall, while requiring 21% fewer parameters. Furthermore, the ensemble maintains real-time inference capabilities, operating at 71 FPS on an NVIDIA T4 GPU, making it well suited for deployment in time-sensitive and safety-critical environments.

A recent manuscript²⁴ using the CTX-UXO dataset has demonstrated effective multi-class identification performance. In that study, the authors introduced a specialized identification pipeline that leverages neuro-symbolic AI together with human-in-the-loop approaches, leading to an end-to-end framework called DEFUSAL, which achieved an F1-score of 85.3%.

In the past, studies were conducted on smaller versions of the dataset, 7,880 instances¹² or 13,340 instances^11,18 to validate intermediate phases, with the CTX-UXO dataset updated regularly as new UXOs were discovered. These studies demonstrate the dataset’s versatility in supporting various approaches to UXO detection, and its potential for deployment in resource-constrained environments. The dataset’s comprehensive coverage of various UXO states and conditions makes it particularly valuable for researchers working on supervised and semi-supervised object recognition projects, with specific applications in civil protection and emergency situation management. By providing a standardised resource for UXO detection research, the CTX-UXO dataset enables the development of more effective solutions for humanitarian demining and emergency response, contributing to the larger goal of building long-term resilience in post-conflict regions.

Methods

Data Acquisition

The CTX-UXO dataset was created through approximately 4 years of data collection process with the help of pyrotechnical personnel from the National Romanian Inspectorate for Emergency Situations. The acquisition protocol was designed to capture UXOs in various states and environmental conditions, exactly as found in the field. The majority of the images were acquired utilizing a mobile phone configuration equipped with a 64MP Samsung GW3 sensor (S5KGW3), characterized by a 1/1.97-inch imager with 0.7 μm pixel dimensions and Tetra-cell technology, in conjunction with a 25mm f/1.8 lens. Each image was captured in RAW format and subsequently converted to JPG while preserving the original aspect ratio and colour space. The dataset encompasses images captured from October 2021 to July 2025, covering all seasons (spring, summer, autumn, and winter) as well as diverse meteorological conditions, including sunny, cloudy, rainy, and snowy weather. Data collection was carried out over various seasons to ensure an extensive representation of diverse environmental conditions. The distribution of data across these seasons and days is non-uniform, accurately mirroring the practical circumstances encountered in real-world UXO clearance operations.

Most of the images were captured in their natural state as found in the field, maintaining the authentic context of discovery - as seen in Fig. 2a. When safety conditions permitted, operators approached the ordnance to capture detailed views, as in Figure 2b. For inert munitions without explosive charges or fuses (replicas or inert), multiple rotations were performed during the training session of the Explosive Ordnance Disposal operators to create various viewing angles and scenarios. As demonstrated by Lebbad et al.¹³, multiple viewing angles and the removal of parts of the unexploded weapon (fuses, tails) can improve the results. When the operation was safe (no explosive charge inside, during the EOD training), the removal of parts method was implemented. Multiple distinct images of the same physical item of ordnance occasionally appear in the dataset under three separate field conditions. The first involves cases in which a UXO is initially recorded in a partially buried condition and is then photographed again after it has been removed. The second concerns controlled demolition activities carried out at specific ranges, during which the ordnance is moved within a managed setting. The third results from the shape of certain munitions, for instance when the large body of an aircraft bomb hides details of the tail fuze, making it necessary to capture separate images of different components in order to achieve comprehensive documentation. The primary objective was to ensure a representative dataset that accurately reflects real-world conditions and the actual operational perspective of UXO detection and identification. Of course, even though the dataset captures multiple hypothetical cases, researchers can reduce its complexity by selecting only those instances that are relevant to their specific problem, e.g. selecting a single class of interest. The main contribution lies, first and foremost, in the acquisition of the dataset, followed by its annotation by a trained Explosive Ordnance Disposal operator and the multiple technical validations.

Fig. 2.

The dataset is composed of images of UXOs found on real pyrotechnical interventions or during training. Pictures were taken in close proximity, in safe conditions.

Regarding the manufacturing origin of the ordnances, they originate from various sources depending on the historical conflicts of Europe, with a predominance of munitions originating from Germany, United States, the United Kingdom, the Soviet Union, Austria-Hungary, France, the Ottoman Empire, Romania and Bulgaria.

Dataset Construction

The annotation process was conducted by a trained UXO expert. Each sample was annotated using a standardised protocol that included the instance mask and classification labels for each instance. The instance mask is resulted from applying Segment Anything Model 2²² over the supervised boundary boxes obtained, with additional refinement when necessary. The annotation process was validated using a human-in-the-loop²⁷ active learning approach, where initial annotations were iteratively refined through multiple cycles of model training, inference, and expert review.

To ensure the highest quality and reliability of the dataset, a comprehensive data quality assurance pipeline was implemented using modern computer vision tools. The dataset was subjected to rigorous quality control using Voxel51’s FiftyOne solution¹⁹, which enabled the analysis of image quality, annotation consistency, and dataset characteristics. The platform facilitated the identification and resolution of potential biases through advanced visualisation and analysis tools. ImageDedup²⁰ was employed for advanced duplicate detection and removal, ensuring dataset diversity while maintaining representative samples of each UXO category. The deduplication process used perceptual hashing algorithms²⁰ to identify images that were nearly duplicated and could introduce bias into the training/fine-tuning process. This deduplication pipeline was executed in two complementary stages and applied to the entire dataset before performing the train-validation-test split²⁸ described in the Data Records section. In the first stage, ImageDedup was run at the full-image level, covering both background and object instances, to identify and remove duplicate and near-duplicate samples using perceptual hashing. In the second stage, FiftyOne’s Brain module was applied at both the instance level and the full-image level to identify and remove visually similar samples using the embeddings. Together, these two stages ensure that images depicting the same piece of ordnance in comparable detail (embeddings) or identical background settings are not spread across the training, validation, and test splits, thereby reducing the risk of data leakage and cross-contamination between partitions.

Preprocessing and Image Enhancement

The detailed and intricate attributes of the dataset allow for the deployment of a diverse range of techniques, as illustrated in recent studies on the detection of unexploded ordnance (UXO)^11,12,18,24. The CTX-UXO dataset can be effectively utilised in combination with various pre-processing strategies that have already been integrated into data curation processes. Furthermore, additional pre-processing methods are recommended for adoption by researchers to enhance robustness when using small feature extractors as the backbone of deep neural networks.

A fundamental step in the preprocessing pipeline was the removal of all Exchangeable Image File Format (EXIF) metadata from each image in the dataset. This procedure specifically addressed the orientation tag, an embedded parameter responsible for indicating the correct rotational alignment of an image as intended by the capturing camera device. The interpretation of this tag is inconsistent across various software platforms, libraries, and viewing applications, potentially leading to unintended rotations and compromising the visual integrity of the data. By removing this attribute, all images were normalised to a standard canonical orientation. This measure ensures dataset uniformity and eliminates a notable source of potential error, thereby guaranteeing that any subsequent analysis or model training is conducted on a visually consistent and standardised set of inputs. Furthermore, to comply with General Data Protection Regulation (GDPR), all sensitive EXIF metadata were completely removed, as they contained information pertaining to the devices employed for image capture, as well as location data. As a result of this sanitisation process, temporal acquisition timestamps and geographic coordinates are not preserved at the individual image level. This choice reflects a zero-risk policy concerning the disclosure of intervention site locations. To report this sanitisation step transparently, an auxiliary file uxos_exif.csv is included in the repository, verifying the absence of such metadata throughout the dataset.

In the past¹² a “copy-paste on new backgrounds”²⁵ method was utilised as a basic augmentation technique and is highly recommended. It is recommended that additional negative samples, empty backgrounds, should be added - as a minimum determined number of 500 images. These samples should include terrain, soil, snow, or any other environment. One such dataset with environments that can be used for augmentation is ADE20K²⁹. Adding such images with backgrounds and using them with copy-paste instance-on-new-background augmentation can improve performance scores for small neural networks¹², especially for instance segmentation task. One challenge faced by the authors of the dataset and other specialists¹ is that munitions are often very well camouflaged in the natural environment. This camouflage may be intentional, to serve a military purpose, or a result of the passage of time. For this reason, a small neural network (e.g. ResNet18 or MobileNet as feature extractor) may tend to learn features unrelated to the camouflaged munition and instead focus on other easy to distinct features. Considering the findings of Yue Wang et al.³⁰, a safety-critical application requires the use of explainable AI (xAI) techniques to validate that the model generalize well because a system failure may have consequences. In figure 3 it is shown a Explainable AI Local Interpretable Model-Agnostic Explanations (LIME) technique²³ used by the authors at the first tests of a smaller version of CTX-UXO to understand why a CTX-UXO classification task failed in specific cases. The formula used for applying LIME is as follows:

$$\xi \left(x\right)=\underset{g\in G}{\arg \min }L\left(f,g,{\pi }_x\right)+\mathrm{\Omega }\left(g\right)$$ 1

Fig. 3.

Explainable AI technique used by the authors at the first tests of a smaller version of CTX-UXO to understand why a classification model may give bad predictions. The munitions are camouflaged, either intentionally for military purposes or due to the passage of time and environmental conditions, making them difficult to identify using classical image processing methods or without specific augmentation techniques.

In the formula above, f denotes the model, x is the instance being explained, G represents the class of interpretable models, g is a candidate interpretable model, L(f, g, π_x) measures the fidelity of g to f in the locality defined by π_x, π_x is a proximity measure around x, and Ω(g) penalises the complexity of g. Based on the authors’ tests and publication¹², copy-paste-on-new-background augmentation addresses this issue effectively. Other classical approaches, without deep learning, based on image processing techniques, such as selective search⁴, may not be suitable in every scenario, considering the complexity of the images. In addition, classical augmentation techniques such as HSV variations, rotation, and translation can be used, with their parameters carefully tuned based on the specific deployment context of the model. However, as demonstrated in study¹¹, the dataset can achieve competitive results even without augmentation techniques when appropriate training strategies, hyperparameter tuning and loss function are implemented. Specifically, the use of the focal loss function²⁶ with the modulating factor γ = 2 and a dynamic learning rate scheduler to reduce LR on the validation loss plateau, with a multiplier factor of 0.5, proved effective in addressing two key challenges: distinguishing between easy and medium camouflaged UXO classes and handling dense instances where multiple UXOs appear in the same sample. The usefulness of focal loss lies in its ability to penalise negative instances, such as background or other classes that are easier to learn, thus helping the model to also learn instances that are harder to classify¹⁸.

Data Records

Figure 4 illustrates visual examples for the classes in the CTX-UXO dataset, openly available in Zenodo, under record number 17052675, at the following link https://zenodo.org/records/17052675³¹.

Fig. 4.

Visual examples for the classes in the CTX-UXO dataset.

Table 1 summarises the file and folder catalogue of the released archive.

Table 1.

File and folder catalogue of the CTX-UXO Zenodo archive.

Path / File	Format	Description
images/train/	JPEG	2461 training images (UXOs_<index>.jpg)
images/val/	JPEG	530 validation images
images/test/	JPEG	529 test images
coco_labels/coco_train.json	COCO JSON	Annotations for training subset
coco_labels/coco_val.json	COCO JSON	Annotations for validation subset
coco_labels/coco_test.json	COCO JSON	Annotations for test subset
yolo_bbox/	YOLO txt + data.yaml	Bounding-box labels, train/val/test split
yolo_segmentation/	YOLO txt + data.yaml	Instance segmentation labels, train/val/test split
seasons.csv	CSV	Per-image season and confidence metadata
instance_sizes.csv	CSV	Per-instance bounding box area and size category

Upon extraction, the archive is organised into the following components: images, coco_labels, yolo_bbox, yolo_segmentation, and two supplementary metadata files, seasons.csv and instance_sizes.csv. The images directory contains the 3520 RGB images in JPEG format, partitioned into three subdirectories, namely train (2461 images), val (530 images) and test (529 images). Each image is named according to the convention UXOs_<index>.jpg. The coco_labels directory contains three JSON files following the standard COCO schema³², corresponding to the respective image subsets. Each file includes the standard COCO fields images, annotations (bounding boxes and instance segmentation polygons) and categories, with category identifiers mapped to the UXO subcategories listed in Table 2. Two parallel YOLO-format repositories are provided to accommodate different task setups. The first, yolo_bbox, contains bounding-box labels for object detection (one .txt label file per image, using the format class_id x_center y_center width height with coordinates normalised to the image size). The second, yolo_segmentation, stores instance segmentation labels (again, one .txt label file per image, where each line begins with class_id followed by normalised polygon vertices). Both yolo_bbox and yolo_segmentation follow the same internal structure, with train, val, and test subfolders for images and labels, and each repository includes its own data.yaml configuration file that defines the class names, the total number of classes, and the corresponding paths to image and label directories. The complete mapping from classes to indices is specified in the data.yaml files in yolo_bbox and yolo_segmentation, and directly matches the categories field in the COCO JSON annotations. Two additional CSV files are provided to support the statistical characterisation of the dataset.

Table 2.

UXO Categories and Instance Counts.

Category	Total Instances
Unexploded Ordnances	15 449
Projectile	6 121
Grenade	4 269
Mortar Bomb	3 399
Cartridge	987
Aviation Bomb	333
Cartridge Magazine	124
Fuse	92
Rocket-Propelled Grenades RPGs	63
LandMines	29
Rockets	21
AntiSubmarine Bomb	6
Sea Mine	5

Unexploded ordnances appear in various states, with single or multiple UXOs depicted within a single image. Although the primary objective of most automated UXO recognition systems remains the binary differentiation between hazardous (UXO) and non-hazardous (non-UXO) entities, the capability to further discern specific UXO types may provide valuable operational insights.

The CTX-UXO dataset incorporates a diverse array of types of ordnance, encompassing projectiles, grenades, mortar bombs, cartridges, aerial bombs, cartridge magazines, fuses, rocket-propelled grenades, landmines, and rockets, each defined by varying risk levels. For example, aerial bombs may feature multiple fuzing mechanisms, such as nose, tail, or lateral fuzes, along with sophisticated activation mechanisms and considerable explosive charge. Despite their compact size, grenades often employ highly sensitive fuzes, making them prone to accidental detonation even with minimal movement. Mortar bombs typically include a stabilisation tail and impact-sensitive fuzes, which require distinct mitigation procedures compared to projectiles. Cartridges or magazines, often seen as low-risk, can still contain propellant under certain conditions. Fuses, attached or isolated, pose significant safety concerns as a result of their role as a detonation initiator. Rocket-propelled grenades and rockets might contain unexploded warheads or reactive propellant elements, while landmines often use pressure-sensitive triggers, complicating the detection and removal processes. Although the prevailing research prioritises binary UXO/non-UXO classification, the type-level annotations available within the CTX-UXO provide potential for facilitating downstream operations such as priority mapping, contextual interpretation, or post-processing strategies. In practical demining scenarios, distinguishing between UXO categories can also help determine appropriate intervention techniques and safety measures, particularly when such systems are mobilised within semi-autonomous or fully autonomous vehicles, including unmanned ground vehicles or aerial vehicles¹¹.

The CTX-UXO imagery dataset is organised into three main repositories (subsets). The Raw Images Repository comprises 3520 high-resolution RGB images, with a median dimension of 2124 pixels squared. These images are divided into training (70%, 2461 samples), validation (15%, 530 samples), and test sets (15%, 529 samples). Considering the presence of multiple instances from various subcategories of unexploded ordnances (UXOs) in the same sample, the dataset was partitioned into training, validation, and test subsets using the Multilabel Stratified Shuffle Split algorithm as introduced by Sechidis et al.²⁸. This approach ensures an equitable representation of all classes across these subsets. To achieve this, each image was encoded as a binary multilabel vector indicating two possible states: the presence or absence of each class. This encoding allowed the algorithm to preserve both the frequencies of individual classes and their co-occurrence patterns. Initially, the instances were divided into a combined training-validation set encompassing 85% of the images, and a test set comprising the remaining 15%, while maintaining the multi-label distribution. Following that, the training-validation set was further subdivided into training (approximately 70%) and validation (approximately 15%) subsets using the same stratification methodology. This partitioning strategy effectively mitigates sampling bias and preserves the comprehensive class distribution and their co-occurrence patterns within each subset, thereby enhancing the development of robust and generalisable models. Table 2 presents the categories of the dataset, accompanied by the number of instances for each category. The first 3 subclasses (mortar bombs, projectiles, grenades) can be used for supervised or semi-supervised tasks, while the other classes should be augmented or used for recognition tasks based on a few shot approach³³ after binary classification (UXO/NON-UXO).

Table 3 presents the distribution of UXO by season, relative to background.

Table 3.

Distribution of UXO by season.

Season	Total	Train (2461)	Valid (530)	Test (529)
Spring	818 (23.2%)	577 (23.4%)	111 (20.9%)	130 (24.6%)
Summer	567 (16.1%)	400 (16.3%)	87 (16.4%)	80 (15.1%)
Autumn	1045 (29.7%)	736 (29.9%)	151 (28.5%)	158 (29.9%)
Winter	235 (6.7%)	172 (7.0%)	34 (6.4%)	29 (5.5%)
Others (Indoor)	855 (24.3%)	576 (23.4%)	147 (27.7%)	132 (25.0%)

In Table 3, a greater occurrence is seen in the warmer seasons, with the number of cases peaking in autumn, likely linked to agricultural activities. It should be clarified that “Others (Indoor)” designates ordnance retrieved from storage sites, concrete-surfaced areas, or locations where the season does not play a significant role.

Table 4 presents the distribution of target size according to dimensions, in accordance with COCO standards³², as follows: small, area  < 32²; medium, 32² < area < 96²; large, area  > 96².

Table 4.

Distribution of target sizes across the dataset splits, in correlation with COCO standards³². The dataset split was performed with respect to the number of UXO classes²⁸.

Target Size	Total	Train (11,146)	Valid (1,862)	Test (2,441)
Small	1,001 (6.5%)	574 (5.1%)	114 (6.1%)	313 (12.8%)
Medium	3,830 (24.8%)	3,155 (28.3%)	297 (16.0%)	378 (15.5%)
Large	10,618 (68.7%)	7,417 (66.5%)	1,451 (77.9%)	1,750 (71.7%)

Technical Validation

For the detection task, the current section is addressing instance localisation and binary classification into UXO/NON-UXO instances. While the dataset supports multi-class identification and instance segmentation tasks through the provided COCO and YOLO format annotations, the technical validation presented in this section focuses specifically on binary localisation and classification, representing the primary operational use case for real-time, resource-constrained deployment scenarios. This approach (UXO/NON-UXO) can be beneficial in cases where munitions have been modified¹² or when there are munitions with rare characteristics, as presented in Fig. 5. For edge cases, based on the scenario when the model should be deployed, a binary classification task can be applied over the resulting region proposals, following other approaches as needed, e.g., few-shot learning for specific identification³³. The current section will include primary and secondary metrics obtained after fine-tuning a Faster-RCNN⁷ model with MobileNetV3³⁴ as the backbone. In the work by³⁴, MobileNetV3 was introduced as a convolutional neural network structure engineered to deliver rapid inference on edge devices. The architecture is derived by integrating inverted residuals from MobileNetV2 with squeeze-and-excitation (SE) blocks, alongside employing an automated neural architecture search (NAS) strategy that targets optimisation for improved processing speed and precision. PyTorch framework was chosen primarily as it is a very popular framework among researchers, secondly to demonstrate how to use the CTX-UXO dataset with custom functions, such as a customisable dataloader.

Fig. 5.

Uncommon types of UXOs may have similar shape with other categories.

For other extended frameworks, such as YOLO, specific annotations in YOLO format are included in the public repositories, both for boundary box and instance segmentation tasks. The Faster-RCNN architecture was chosen because it is a two-step object detection architecture, addressing localisation and classification sequentially, thereby offering good precision. In Fig. 6, qualitative samples from the CTX-UXO test set are illustrated after the inference with the fine-tuned Faster-RCNN.

Fig. 6.

Qualitative samples resulted from Faster R-CNN MobileNet fine-tuned on CTX-UXO dataset.

Following the paradigm already presented by the authors, a UXO detection model must respond not only to achieving good primary metrics and a low false negative rate, but also to reduced computational resource consumption and appropriate inference speed. This ensures that the solution responds both to the relevance of not limiting lifesaving solutions in terms of costs¹², and maintains the real-time detection paradigm, in the event that the solution is implemented in embedded devices¹¹, such as unmanned vehicles. The scenario is a pessimistic one, as the Nvidia Jetson has limited resources but offers portability, and the two-step detector requires additional optimisation for the real-time paradigm. This optimisation is performed using kernel auto-tuning and layer fusion with ONNX Runtime and TensorRT frameworks. At the same time, the model is postquantised from FP32 to FP16. The methodology for exploring the real-time paradigm using TensorRT and FP16 has already been tested and validated by other specialised works that investigate similar capabilities, such as You Only Look Once¹⁰ (YOLO) and Real-Time Detection Transformer³⁵ (RT-DETR). In Fig. 7 the evolution of the precision measured for a small Faster R-CNN MobileNetV3, with 5.4 million parameters for the feature extractor, is represented. In Table 5, the results obtained on the test subset are shown. The detailed metrics comprise the findings from the implementation of the optimised model on the Nvidia Jetson Orin platform, equipped with 16GB of VRAM. To have a real reference, the primary metrics are standardised³² and include details regarding precision at different Intersection over Union (IoU) thresholds, as well as recall. For secondary metrics, inference speed is expressed through conversion to processed Frames per Second - FPS. As a reference, the same architecture has an mAp@0.5:0.95 of 32.8% on COCO-Val2017.

Fig. 7.

Evaluation metrics on CTX-UXO Dataset over training-validation epochs. With proper hyperparameter tuning and augmentation methods, accessible models can be used for real-world UXO detection.

Table 5.

Performance metrics on test subset of the model deployed on Jetson Orin for real-time UXO detection.

Perfomance Metric	Value
mAP@0.5	62.18%
mAP@0.75	42.49%
mAP@0.5:0.95	39.57%
mAR@100	43.66%
FPS	23

For comparison, other architectures have been tested on the CTX-UXO dataset, including You Only Look Once (YOLO) V8¹², RetinaNet¹¹, the Real-Time Detection Transformer (RT-DETR)¹⁸, and neuro-symbolic AI with multimodal capabilities for specific target identification²⁴.These works are discussed in greater detail in the Section Background and Summary.

The available code repository includes an automated validation pipeline that checks the dataset at several stages before any training or evaluation takes place. At the structural level, a sanity-check module programmatically verifies that the required directory layout (train/, val/, test/) exists and that the data.yaml manifest file is present. When the dataset is obtained using the supplied download scripts (Zenodo API), the pipeline automatically extracts the archives, restructures the folders to combine image and annotation directories into the required split-based format, and re-checks for the manifest file after extraction, returning an explicit failure signal if it is missing. At the per-sample level, the custom dataloader attempts to load each image with OpenCV and confirms that decoding succeeds. Any image that cannot be read (for instance, because it is corrupted or truncated) is skipped automatically with a warning recorded in the logs, and the loader proceeds to the next valid entry. If, after iterating over all candidates, no images can be loaded, a runtime error is thrown to avoid silent data omissions. All successfully loaded images are converted from BGR to RGB colour space to align with the normalisation pipeline.

In this section, we fine-tuned a Faster R-CNN model to demonstrate that the dataset remains valid and qualitative even for edge-deployment scenarios, which represent a realistic real-world setting (e.g., UGVs, UAVs).

Usage Notes

The CTX-UXO dataset is suitable for object detection, instance segmentation, and classification of unexploded ordnance. Researchers focusing on instance segmentation or multi-class detection should refer to the modular code repository, which includes appropriate dataloaders and training scripts for these extended tasks. The dataset accommodates several task formulations. The dataset also enables multi-class identification and instance segmentation through its COCO and YOLO-format annotations. For YOLO-based detection and segmentation pipelines, dedicated YOLO-format annotations are provided in the public repositories for both bounding box and instance segmentation tasks. Researchers working on these extended scenarios are encouraged to use the available modular code repository, which offers the corresponding dataloaders and training scripts.

Drawing on previously validated work^11,12,18, researchers are recommended to fine-tune single-stage detectors pre-trained on the Common Objects in Context (COCO) dataset³². For two-stage architectures, it is advisable to employ feature extraction backbones initially trained on ImageNet³⁶. A UXO detection model should be optimised not only for strong primary performance metrics and a low false negative rate, but also for low computational overhead and suitable inference speed. This balance ensures that the solution remains affordable and does not restrict the deployment of life-saving systems¹², while preserving the real-time detection capability necessary for embedded platforms¹¹, such as unmanned vehicles.

Users must comply with all relevant safety procedures when operating with real UXOs, as any UXO constitutes a serious hazard. As a general safety principle, a human-in-the-loop should be maintained: the final confirmation of UXO presence and its precise identification must be carried out by trained personnel, with the role of the model(s) being to automate and substantially ease the process, offering multi-layered support when the operator is uncertain about the type or even the existence of a UXO. Users are also encouraged to adopt the provided preprocessing pipeline and adhere to the recommended train/validation/test splits.

Author contributions

M.C. conceived the data collection protocol, data labelling, and data curation. He also developed the training code and he deployed the models on the Nvidia Jetson Device. He contributed in writing the manuscript. C.V. conceptualised the manuscript, tuned hyperparameters, and developed the inference scripts for the Validation Section. He also contributed to the writing of the manuscript. G.S. designed the active learning annotation process, developed a preprocessing pipeline, optimisation techniques, and supervised the project. He also organised the structure of the project. All authors reviewed and approved the manuscript.

Data availability

The dataset that represents the main object of this study is openly available in Zenodo, under record number 17052675, at the following link https://zenodo.org/records/17052675.

Code availability

The code for dataset preprocessing, validation and analysis is available at https://github.com/grig101/ctx-uxo. The repository includes preprocessing scripts, training scripts, validation tools, metrics visualisation, and inference optimisation techniques. The code is provided under the MIT licence and includes detailed documentation for installation and usage. The installation of dependencies is included in the repository, and the example scripts demonstrate the complete workflow from data loading to model training and evaluation. As a supplementary help for researchers or operators, a containerised version with Application Programming Interface of the end-to-end pipeline exposed can be found in the same repository.

For annotation checking, the dataloader reads each YOLO-format label file, extracts polygon vertices, converts them into bounding boxes in absolute pixel coordinates, and clamps these coordinates to the valid image domain [0, S], where S denotes the input resolution. This clamping step prevents invalid coordinates from propagating into the training or evaluation procedures. Moreover, during COCO-style evaluation, both predicted and ground-truth bounding boxes are filtered to remove degenerate instances where $x_{\min }\ge x_{\max }$ or $y_{\min }\ge y_{\max }$, as well as any box with zero or negative area. For the binary detection setup, all original subcategory labels are collapsed into a single positive class (UXO), and images without any annotations are represented via empty tensor initialisation, ensuring that the pipeline robustly handles both positive and negative examples.

Competing interests

The authors declare no competing interests.