InspectGaussian: Large-scale coarse-to-fine Gaussian reconstruction for orchard inspection robots

Abstract

Efficient large-scale 3D reconstruction of orchard environments is essential for robotic inspection and precision agriculture, yet existing methods struggle with unstructured scenes, variable illumination, and computational bottlenecks. We propose InspectGaussian, a coarse-to-fine Gaussian reconstruction framework tailored for orchard inspection robots. The pipeline integrates an RGB-D-based data acquisition strategy using ORB-SLAM3, which is enhanced by a dense mapping module for robust large-scale pose estimation and point cloud generation. A divide-and-conquer strategy is then employed: individual plant views are extracted via a YOLO-World–based detection and 3D matching algorithm, followed by plant-specific reconstruction using an improved 3D Gaussian Splatting (3DGS) method incorporating depth regularization and region-aware refinement. Experimental results in citrus orchards demonstrate that InspectGaussian achieves 96% average precision and 93% recall in plant view extraction, while surpassing state-of-the-art methods in reconstruction fidelity (31.226 PSNR, 0.915 SSIM, 0.067 LPIPS) and point cloud accuracy (7 mm error). These results confirm its effectiveness in capturing fine structural and textural details while maintaining scalability and efficiency. This framework provides a practical solution for high-throughput, in-field plant phenotyping and lays the foundation for intelligent orchard monitoring and management.

1. Introduction

Citrus, accounting for over 20% of global fruit production, is one of the world's most important fruit crops. Accurate analysis of phenotypic traits, including tree architecture, fruit appearance, and responses to pests and diseases, is essential for efficient orchard management and high fruit quality [1,2]. Traditional manual measurements, however, are time-consuming and prone to bias, making them insufficient for the demands of precision agriculture and smart orchard management, which require large-scale and high-throughput data acquisition [3,4].

Three-dimensional reconstruction has emerged as a key technology for automated phenotyping because it captures detailed structural information needed for trait extraction and dynamic monitoring. However, applying 3D reconstruction at orchard scale remains challenging. Typical data-collection workflows require manual circumvention of individual trees, which is labor-intensive and disruptive; high planting density and complex, unstructured plant geometry further reduce efficiency and data quality. Outdoor conditions introduce additional difficulties: wide variations in plant size, rapidly changing illumination, and the absence of geometric features easily degrade reconstruction robustness.

A range of reconstruction approaches has been developed to mitigate these issues. Early hardware-centric scanners produced 3D models with high efficiency but were constrained by high equipment cost and limited accuracy [[5], [6], [7], [8], [9]]. Photogrammetry lowered sensor cost and enabled UAV-based large-area mapping from RGB imagery, yet performance remains sensitive to image quality and lighting, often falling short for fine-grained phenotyping [10,11]. SLAM and Structure from Motion (SfM) pipelines improved robustness to noisy inputs and complex scenes by coupling camera pose estimation with point-cloud registration, but these gains come with higher computational cost and longer processing time [[12], [13], [14], [15], [16], [17]]. More recently, neural rendering and multi-view learning methods have delivered enhanced reconstruction fidelity and realism [[18], [19], [20], [21]], yet their substantial memory and processing demands hinder large-scale, in-field deployment.

These limitations motivate the need for a reconstruction framework that is both robust to the environmental and morphological variability of orchards and practical for deployment on autonomous data-collection platforms. This study proposes InspectGaussian, a large-scale coarse-to-fine gaussian reconstruction pipeline applicable to orchard inspection robots. By combining deep learning, SLAM, and novel view generation algorithms, low-cost yet efficient orchard 3D reconstruction can be realized. Data collection is achieved using an RGB-D camera fixed on an unmanned ground vehicle. We then use ORB-SLAM3 to construct large-scale orchard maps and employ YOLO-World to extract individual plant viewpoints. Single plant models are finally reconstructed using improved 3DGS to optimize plant details in the orchard map.

The main contributions of this study include:

• •Data acquisition enabled by the proposed pipeline following perimeter shooting along tree ridges rather than circulating each plant, which supports efficient, robot-oriented multi-view capture in scenes with a high planting density;
• •Accurate large-scale point cloud generation through RGB-D based SLAM and a customized dense mapping module, effectively mitigating illumination variability and large inter-plant scale differences;
• •An optimized 3DGS method incorporating depth regularization and region-aware refinement for single plant reconstruction, balancing reconstruction fidelity with computational efficiency to enable practical, large-scale phenotyping.

2. Related work

2.1. 3D reconstruction of plants

In orchards, high-quality point clouds are essential for accurate phenotyping. A major part of reconstruction methods is broadly active using LiDAR, terrestrial laser scanning, and structured-light, offering direct depth at high cost and with limited coverage due to occlusion and viewpoint constraints [[22], [23], [24], [25]]. Other techniques are generally passive based on Multi-View Stereo Network (MVS), Signed Distance Function (SDF), Neural Radiance Fields (NeRF), and 3DGS, offering low-cost, image-driven geometry but relying heavily on image quality and texture [[26], [27], [28], [29]]. Orchard-specific factors, including variable illumination, sparse texture, dense occlusion, and large inter-plant scale variation, typically complicate active scanning and degrade passive reconstructions. Therefore, scalable orchard phenotyping demands reconstruction pipelines that balance robustness to environmental and morphological variability with computational and deployment efficiency.

2.2. 3D reconstruction based on SLAM

Advances in sensor hardware and SLAM algorithms have enabled new multi-scale 3D reconstruction solutions for both indoor and outdoor environments. Depending on task requirements, systems employ monocular, stereo, RGB-D, or LiDAR sensors, often combined with IMU and multi-sensor fusion, to trade off cost, coverage, and robustness. Real-time vision-based SLAM frameworks such as ORB-SLAM (monocular) [30], ORB-SLAM2 (stereo/RGB-D) [31], and visual-inertial systems like VINS-Mono and ORB-SLAM3 [32,33] provide accurate online pose estimates and sparse maps. Compared with direct methods such as DSO, ORB-SLAM3 supports seamless map merging and maintains stable trajectory estimation under prolonged degraded visual conditions through place recognition and loop closure mechanisms [34]. In contrast to VINS-Mono, which adopts a loosely coupled visual–inertial formulation without multi-session mapping, ORB-SLAM3 enables unified multi-map management and faster IMU initialization [35]. However, ORB-SLAM3 focuses on sparse feature-based mapping and does not directly generate dense, colorized point clouds, motivating the integration of additional multi-view or dense reconstruction techniques.

2.3. 3D reconstruction with novel view synthesis

In large-scale orchard scenes, point clouds obtained from traditional Structure-from-Motion or Multi-View Stereo pipelines often lack the density and accuracy required for high-precision plant phenotyping, which has motivated the development of more expressive 3D representation methods. NeRF excels at high-quality view synthesis and modeling complex appearance by optimizing a volumetric radiance field, but their rendering fidelity and training time scale with Multi-Layer Perceptron (MLP) size and optimization cost [28]. By contrast, 3DGS represents scenes with explicit Gaussian primitives, removing heavy MLP dependence and improving spatial fidelity and training speed [36]. However, field environments, marked by severe occlusion, variable illumination, sparse texture, and large scale variations, complicate both approaches.

Our work combines the rich appearance modeling of neural rendering with the efficiency and explicit geometry of point-primitive representations, together with task-aware regularization, offering a solution to obtain dense, colorized, and computationally practical reconstructions for detailed phenotypic analysis in large scale.

3. Methods

The proposed pipeline consists of four stages, as shown in Fig. 1. First, an autonomous ground vehicle equipped with an RGB-D camera acquires synchronized citrus plant images and depth maps throughout the orchard. Second, ORB-SLAM3 is employed for camera pose estimation and initial sparse point cloud construction. Third, an ID-matching module identifies and extracts plant images associated with specific targets. Finally, an improved 3DGS framework generates individual plant models, which are then merged to produce a large-scale orchard reconstruction.

Fig. 1.

Orchard reconstruction framework. A. Data acquisition based on inspection robots and RGB-D images. B. Pose estimation based on SLAM. C. Individual plant viewpoint extraction using 2D and 3D matching. D. Refined single plant reconstructions with 3DGS, producing the final large-scale orchard map.

3.1. Data acquisition

The dataset was collected using a customised autonomous vehicle equipped with LiDAR, an IMU, and an RGB-D camera [37]. The LiDAR and IMU modules facilitate path planning and localization, enabling the vehicle to autonomously navigate citrus rows and perform complete looped data acquisition. Unlike traditional approaches that manually circle individual plants (Fig. 2A), our row-based collection strategy (Fig. 2B) enables smoother and more efficient data capture while minimizing mechanical interference with plants. To ensure uniqueness and traceability, QR code markers with plant-specific identifiers were placed around each plant and decoded from the captured images, thereby effectively preventing cross-contamination of RGB-D data between plants. Collectively, this system provides high-quality input for subsequent 3D reconstruction tasks.

Fig. 2.

Comparison of different data acquisition schemes. A. Conventional method via manually circling each citrus plant. B. The proposed scheme of circling between ridges using an inspection robot.

3.2. Pose estimation

ORB-SLAM3 is a state-of-the-art framework that integrates a multi-map system (Atlas) for enhanced robustness and accuracy across various sensor configurations. In this study, we employ ORB-SLAM3 to recover camera poses from RGB-D images and extend it with a dense module to enhance point cloud quality. The framework comprises six core components: tracking, local mapping, loop closure and map merging, global optimization, dense reconstruction, and map management. The tracking module estimates the camera pose for each frame by extracting ORB features and matching them against a local map; the local mapping module processes new keyframes to perform local bundle adjustment, refining the 3D map points and poses of nearby keyframes; the loop closing and map merging module detects revisits to previously seen areas, correcting drift by computing a similarity transformation and performing pose-graph optimization; the global optimization module executes a full bundle adjustment to minimize global reprojection errors after loop closures; and the map management system handles multiple submaps, activating a new map upon tracking loss and facilitating map merging upon place recognition. Our dense reconstruction module generates colorized dense point clouds, which are jointly optimized across all keyframes. Each keyframe pose is parameterised by a quaternion q for rotation and a translation vector $\mathbf{t}={[t_x,t_y,t_z]}^T$. By transforming and integrating the depth maps of all keyframes, a globally consistent and densely reconstructed 3D point cloud of the orchard is produced. The keyframe pose q is formulated as

$$\mathbf{q}=q_w+q_{x}i+q_{y}j+q_{z}k,$$ (1)

where i, j, and k satisfy the following properties:

$$\left\{\begin{array}{ll}{i}^2=j^2=k^2=-1& \\ ij=k,& ji=-k\\ jk=i,& kj=-i\\ ki=j,& ik=-j.\end{array}\right)$$ (2)

To efficiently compute keyframe transformations, the quaternion obtained from pose estimation is converted into a rotation matrix R ∈ SO(3) as

$$R=\left[\begin{array}{lll}1-2q_y^2-2q_z^2& 2q_x{q}_y-2q_w{q}_z& 2q_x{q}_z+2q_w{q}_y\\ 2q_x{q}_y+2q_w{q}_z& 1-2q_x^2-2q_z^2& 2q_y{q}_z-2q_w{q}_x\\ 2q_x{q}_z-2q_w{q}_y& 2q_y{q}_z+2q_w{q}_x& 1-2q_x^2-2q_y^2\end{array}\right].$$ (3)

To align depth information and point cloud within a global coordinate system anchored to the initial keyframe, point cloud transformation and coordinate unification across all keyframe depth maps are then enabled using

$$\left[\begin{array}{l}{X}_w\\ Y_w\\ Z_w\end{array}\right]=R\left[\begin{array}{l}\frac{(u-c_x)}{f_x}\cdot d\\ \frac{(v-c_y)}{f_y}\cdot d\\ d\end{array}\right]+\left[\begin{array}{l}{t}_x\\ t_y\\ t_z\end{array}\right],$$ (4)

where f_x and f_y denote the focal lengths (in pixels), c_x and c_y represent the principal point coordinates, u and v indicate the pixel locations, and d corresponds to the depth value for each pixel. As a result, a densely reconstructed and globally consistent 3D point cloud of the orchard is achieved.

3.3. Single plant view extraction

Recent advancements in 3DGS [36] have significantly advanced 3D reconstruction. However, as scene scales expand and image sequences increase, the training and rendering processes of 3DGS place substantially higher demands on GPU memory, which in turn degrades reconstruction quality. To address this challenge, we introduce a divide-and-conquer strategy. Instead of reconstructing all plants in a row simultaneously, we perform sequential reconstructions of individual plants. To enable 3D reconstruction of individual plants, a viewpoint extraction algorithm is developed that integrates 2D object detection with 3D spatial reasoning, as shown in Fig. 3. For the detection of plants and QR code markers in RGB images, we employ YOLO-World [38]. Our implementation follows its ”prompt-then-detect” paradigm: we first define a set of descriptive text prompts, namely [QR Code card towards right, QR Code card towards left, the tree in the closest row]. The model's Transformer-based text encoder processes these prompts into textual embeddings, creating a customized offline vocabulary. This vocabulary is subsequently integrated into the RepVL-PAN module, where it undergoes cross-modal fusion with the multi-scale visual features extracted by the YOLOv8 backbone. This process allows for precise identification of the target objects within the images. By pre-computing the offline vocabulary, the model achieves efficient inference as the text encoder is no longer required during the detection phase. After detection, we obtain both the orientation of QR code markers and the positions of plants within the images. Based on the marker orientation, we establish the spatial correspondence between each marker and its nearby plant: markers oriented to the left are associated with the plant on their right, while markers oriented to the right are associated with the plant on their left. This orientation-based reasoning enables an initial 2D matching between plants and markers. Although this 2D matching procedure offers flexibility across diverse scenarios, its performance is constrained by detection accuracy and image resolution, as each frame is processed independently [39]. To overcome this limitation, depth maps and camera poses estimated by ORB-SLAM3 are incorporated. Bounding box contents from keyframes are projected into 3D space to establish prior spatial knowledge, enabling reprojection of plants with failed 2D matches into the world coordinate system, where correspondences are verified against this prior knowledge. This integration ensures robust 3D matching and allows accurate extraction of image sequences for each individual plant, providing a reliable foundation for high-quality reconstruction.

Fig. 3.

Single plant view extraction framework. The first step (yellow background) shows 2D matching based on YOLO-World using RGB images. The second step (blue background) represents 3D matching, achieved by integrating 2D matching results with 3D spatial information. Finally, the third step (purple background) illustrates the extraction of results from the perspective of individual plants.

3.4. An improved 3DGS framework

The complexity of backgrounds and surrounding environmental structures often leads to redundant splitting and duplication of Gaussian primitives when modeling background geometry, thereby diverting reconstruction capacity away from the target plant regions. To mitigate this issue, we propose a regional reconstruction approach. Pixels corresponding to target plant regions are assigned a value of 1, while background areas are set to 0. This mask-weighted strategy is incorporated into the original loss function to emphasise reconstruction accuracy in target plant regions while suppressing background interference, which is formulated as

$$L_{\text{mask}}=(1-{\lambda }_{\text{dssim}})\cdot L_1^m+{\lambda }_{\text{dssim}}\cdot \left(1-{\text{SSIM}}_{\text{mask}}\right),$$ (5)

where λ_dssim denotes the weighting parameter for the DSSIM loss, and $L_1^m$ represents the L1 loss function, defined as

$$L_1^m=\frac{1}{N_{\mathit{mask}}}\sum _{i,j}M(i,j)\cdot \left|I_{\text{render}}(i,j)-I_{\text{gt}}(i,j)\right|,$$ (6)

where I_render(i, j) and I_gt(i, j) are the pixel values at the corresponding positions in the rendered image and the ground truth image, respectively. The structural similarity index (SSIM) is defined as

$${\text{SSIM}}_{\text{mask}}=\frac{(2{\mu }_I{\mu }_{I_{\text{gt}}}+C_1)(2{\sigma }_{I{I}_{\text{gt}}}+C_2)}{({\mu }_I^2+{\mu }_{I_{\text{gt}}}^2+C_1)({\sigma }_I^2+{\sigma }_{I_{\text{gt}}}^2+C_2)},$$ (7)

where C₁ and C₂ are stabilization constants for the formula, μ_I, ${\mu }_{I_{\text{gt}}}$, σ_I, and ${\sigma }_{I_{\text{gt}}}$ represent the pixel mean and standard deviation of the rendered image and ground truth image, respectively, and ${\sigma }_{I{I}_{\text{gt}}}$ denotes their covariance.

Some existing methods employ monocular depth estimation models [40,41] to predict depth from RGB images, using these predictions to supervise depth loss during the early stages of Gaussian model optimization. However, these methods often produce depth maps with inconsistent scales across images because they fail to account for spatial relationships between image sequences, thereby inheriting scale ambiguity. To overcome this, we employ depth cameras to capture depth data with a unified scale, thereby rapidly adjusting primitives to appropriate positions during the initial optimization phase and facilitating more detailed subsequent optimization. As shown in Fig. 4, we incorporate a depth regularization module into several key components of 3DGS, which adjusts the position μ and opacity α of primitives within the Gaussian distribution. To avoid adversely affecting color rendering, the rotation parameters q and scaling parameter s representing covariance are frozen during this process. This is critical because the anisotropic shape and orientation defined by q and s are intrinsically linked to the view-dependent color appearance modeled by the spherical harmonics (SH) coefficients. Freezing them prevents the depth supervisor, which provides a geometrically smooth but texturally coarse signal, from distorting the finely tuned Gaussian shapes that are essential for rendering high-frequency color details and textures. Thus, by updating only μ and α to correct the scene's geometry, while preserving the pre-optimized shape parameters (q, s), we effectively decouple geometric correction from color appearance, safeguarding the rendering quality. We employ a method that combines hard depth and soft depth for depth regularization. Hard depth regularization fixes the opacity α of Gaussian primitives while adjusting their positions μ, aligning the rendered hard depth with real depth data from depth cameras. The rendered hard depth map D_hard is composed of the nearest Gaussian primitives along rays projected from the camera center through each pixel, with the opacity of all primitives set to a high constant value before rendering. The hard depth loss L_hard is calculated as

$$L_{\text{hard}}=L_2({\mathcal{D}}_{\text{hard}},\tilde{\mathcal{D}}),$$ (8)

where $\tilde{\mathcal{D}}$ denotes the depth map from the sensor. Adjusting only the primitive positions often leads to an excess of semi-transparent Gaussians; therefore, we adopt a soft depth regularization approach to further refine their opacity α. At this stage, the positions μof the primitives are fixed, the depth map is rendered to obtain the soft depth map D_soft, and the soft depth loss L_soft is calculated as

$$L_{\text{soft}}=L_2({\mathcal{D}}_{\text{soft}},\tilde{\mathcal{D}}).$$ (9)

Fig. 4.

3DGS with region-aware reconstruction and depth regularization.

The combined hard-soft depth loss is calculated as

$$L_{\text{depth}}=L_{\text{hard}}+L_{\text{soft}}.$$ (10)

4. Results

4.1. Dataset information

To ensure image clarity and accurate alignment between RGB images and depth maps, data were acquired at a resolution of 1280 × 720 and 30 fps. We collected a total of two separate batches of datasets. The first batch was primarily used for the validation of individual algorithmic modules, comprising three datasets: Wuhu-1 (10 plants, 1197 images), Wuhu-2 (10 plants, 1454 images), and Wuhu-3 (10 plants, 1178 images). The second batch was employed to evaluate the effectiveness of the framework in practical orchard phenotypic measurements, including Hzau-2 (7 plants, 1166 images), Hzau-3 (9 plants, 1436 images), Hzau-4 (9 plants, 1559 images), and Hzau-5 (7 plants, 783 images). For the second batch, manual measurements of plant height and primary branch angles were recorded for 32 individual plants, and high-precision point cloud data were acquired using a Livox Mid-360 LiDAR sensor. All data were collected at the same time point to ensure temporal consistency and measurement reliability. For training and validating deep learning–based object detection models for plant and QR code marker recognition, the dataset was annotated. To reduce labeling costs, frames were further subsampled at 2 fps and bounding boxes were annotated using LabelMe software. The experimental hardware environment is summarized in Table S1, which can be found in the Supplementary Material.

Experimental validation employed a three-fold cross-validation strategy, where one sequence was used for training and the remaining two for testing in each trial, and final metrics were averaged to ensure robustness. To assess the performance of the matching algorithm among plant views, markers, and IDs, we manually annotated ID correspondences for each plant across all images, thereby constructing ground truth sets of image filenames for each plant ID. For plant reconstruction validation, 10% of images per plant ID were randomly sampled, and ground truth references were generated by overlaying detected bounding box masks onto the original frames. This protocol ensures a rigorous and objective evaluation of reconstruction fidelity.

4.2. Pose estimation experiment

We employed ORB-SLAM3 to estimate camera poses in orchard scenes and benchmarked its performance against a standard SfM algorithm, COLMAP [42]. We compared the Effective Frame Ratio and processing time of the two algorithms, where the Effective Frame Ratio refers to the proportion of successfully tracked frames within the trajectory. In comparative experiments, SfM failed completely when given the full image sequence, producing fewer than five valid frames across all three datasets. Only after reducing each sequence to 500 images did COLMAP achieve more than 75% valid frames. To further assess ORB-SLAM3, we measured runtime and pose estimation success rate. Results summarized in Table 1 show that ORB-SLAM3 consistently outperforms SfM in both metrics: it operates in real time across all datasets and estimates poses for every frame. In contrast, SfM requires substantially longer runtime on the full sequences and achieves very low success rates, highlighting its limitations in large-scale scenarios. Even with reduced inputs, its runtime remains far higher than ORB-SLAM3 and it fails to recover reliable poses for all frames. These findings clearly demonstrate the superior efficiency and robustness of ORB-SLAM3. Furthermore, we provide a comparative visualization of the dense mapping module integrated into ORB-SLAM3, with detailed results presented in Fig. S1.

Table 1.

Comparison of pose estimation algorithm performance.

Metric	Dataset	ORB-SLAM3	SfM (500 Frames)	SfM (All image sequence)
Time (minutes)	Wuhu-1	1.995	92.577	548.329
	Wuhu-2	2.423	117.41	738.132
	Wuhu-3	1.963	109.552	1091.388
Effective frame ratio	Wuhu-1	1.000	0.794	0.003
	Wuhu-2	1.000	1.000	0.003
	Wuhu-3	1.000	0.992	0.003

We extended ORB-SLAM3 with a dense reconstruction module to build a large-scale orchard map. During reconstruction, RGB-D images were converted to point clouds with a 3 m depth threshold to suppress noise, followed by additional denoising. As shown in Fig. S1, the reconstructed point clouds are generally clean, with clear plant outlines and well-defined QR code markers. However, noticeable drift remains between plant and marker point clouds, likely due to limited depth sensor accuracy.

4.3. Single plant view extraction experiment

We evaluated the YOLO-World model in complex environments with specialized textual contexts and limited data, comparing it against state-of-the-art (SOTA) models including YOLOv8, YOLOv10, YOLOv11 [43], and Detection Transformer(DETR) [44]. As shown in Fig. 5A, YOLO-World achieved precision, recall, and F1-scores of 0.860, 0.846, and 0.850, respectively, outperforming the other models and reflecting low false-positive and false-negative rates. We further assessed its zero-shot detection capability: without fine-tuning, YOLO-World reached 94.9% precision and 7.5% recall, while the other models failed to detect targets under these challenging conditions.

Fig. 5.

Results of single plant view extraction experiments. A. Mean performance across three-fold cross-validation of object detection models. B. Comparison of the performance of 2D and 3D matching methods. ∗ indicate statistically significant differences. C. Visual results of the target detection model. Areas highlighted in red indicate detection errors.

To quantitatively assess the effectiveness and robustness of the proposed 3D matching algorithm, we conducted an ablation study against its 2D counterpart. Using the YOLO-World model trained on the Wuhu-1 image sequences, we evaluated the matching among plants, QR markers, and IDs on the Wuhu-2 and Wuhu-3. As shown in Fig. 2, Fig. 5D matching achieved marginally higher precision but suffered from markedly lower recall, with a maximum of only 84%. In contrast, 3D matching consistently achieved higher F1-scores, indicating better balance between precision and recall. These results demonstrate that by exploiting spatial information from camera poses, the 3D matching algorithm provides more robust and generalizable results.

To illustrate YOLO-World's advantages, we visualized sample detections highlighting three common error types, including missed detection, false detection of objects in the current row, and false detection of objects in other rows, as shown in Fig. 5C. The most frequent error was incorrect orientation of QR code markers (false detections), highlighting the model's sensitivity to local contextual textual cues. This capability enables YOLO-World to effectively leverage textual information for accurate object detection and contextually appropriate labeling.

4.4. Orchard reconstruction experiment

Building on single plant view extraction, we performed global reconstruction experiments using 10% of images per plant ID as the test set. As shown in Table 2, the proposed divide-and-conquer plant reconstruction algorithm substantially outperformed competing methods on both the Wuhu-2 and Wuhu-3. It achieved mean values of 31.226 (PSNR), 0.915 (SSIM), and 0.067 (LPIPS), surpassing original 3DGS, SplaTAM, and Photo-SLAM. These results highlight its advantages in pixel-level detail restoration, more accurate fine-structure capture, and enhanced spatial consistency.

Table 2.

Comparison of reconstruction quality.

Method	Modules		Wuhu-2			Wuhu-3
	Mask	Depth	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
InspectGaussian (this study)	✓	✓	30.518	0.903	0.074	31.933	0.927	0.059
	✘	✘	30.362	0.903	0.079	31.500	0.920	0.066
	✓	✘	30.372	0.899	0.077	31.636	0.920	0.064
	✘	✓	30.456	0.904	0.075	31.531	0.922	0.065
3DGS [36]	✘	✘	23.912	0.657	0.409	23.245	0.673	0.398
SplaTAM [45]	✘	✓	17.479	0.530	0.431	15.198	0.536	0.395
Photo-SLAM [46]	✘	✓	11.705	0.474	0.657	5.048	0.006	0.724

To evaluate the contributions of different improvement modules to single plant reconstruction, we compared performance across various module combinations, all under the divide-and-conquer strategy where plants for each ID were reconstructed sequentially. As shown in Table 2, integrating all modules yielded the best overall performance, whereas the baseline without improvements performed the worst. Notably, on the second row, the PSNR of the depth-only module reached 30.456, nearly matching the 30.518 achieved by the full method. This highlights the critical role of ground-truth depth in enhancing reconstruction quality, as it effectively regularizes geometric optimization and reduces ambiguity in multi-view consistency.

To visually compare the reconstruction quality of different methods, we rendered the Gaussian models generated by various algorithms on the test set, with the results presented in Fig. 6. As shown in the figure, our proposed method achieves visual results most consistent with the ground-truth images across all plants and under varying lighting conditions. The SplaTAM method demonstrates poor robustness in pose estimation and mapping, producing chaotic rendering results that fail to capture plant structures. While the original 3DGS can reconstruct plants using the pose and point cloud initialization provided in this study, it introduces pronounced artifacts and blurring, particularly along leaf edges and in densely foliated regions, thus limiting its applicability to high-precision tasks such as plant phenotyping.

Fig. 6.

Visualization of reconstructions from different methods. The red boxes indicate the differences in detail between ours and 3DGS.

To comprehensively evaluate the quality of reconstructed point clouds in 3D space, we employ the accuracy and completeness metrics proposed by Ref. [47] for quantitative evaluation. Accuracy is measured as the mean distance from each point in the reconstructed point cloud to its nearest neighbour in the ground-truth point cloud, whereas completeness is the mean distance from each point in the ground-truth point cloud to its nearest neighbour in the reconstruction. The overall metric is calculated as the mean of accuracy and completeness, providing a balanced measure that simultaneously accounts for positional accuracy and reconstruction coverage. Together, these complementary metrics evaluate reconstruction fidelity in terms of both precision and coverage. We project the captured ground-truth depth onto the real point cloud, which is then used as the reference for evaluating the reconstructed point cloud. As shown in Table 3, our method achieves lower errors in both accuracy and completeness compared to the other three approaches. Notably, SplaTAM fails to generate valid point clouds on the Wuhu-2 dataset, producing results that consist almost entirely of outliers. This suggests that as the number of input images increases, global reconstruction-based methods are prone to point cloud drift and often fail to achieve accurate and stable reconstructions.

Table 3.

Evaluation of reconstructed point cloud quality across different datasets.

Data	Metrics (mm)	InspectGaussian	3DGS	SplaTAM	Photo-SLAM
Wuhu-2	Accuracy	7.33	11.11	—	11.59
	Completeness	7.02	12.80	—	12.94
	Overall	7.17	11.95	—	12.265
Wuhu-3	Accuracy	7.20	10.86	12.36	9.89
	Completeness	7.25	12.96	13.63	12.76
	Overall	7.22	11.91	12.99	11.33
Hzau-2	Accuracy	12.74	12.82	13.94	12.85
	Completeness	7.82	12.50	13.56	12.43
	Overall	10.28	12.66	13.75	12.64
Hzau-3	Accuracy	11.10	10.91	11.30	11.98
	Completeness	7.68	12.98	13.09	13.73
	Overall	9.39	11.95	12.20	12.85
Hzau-4	Accuracy	11.53	11.23	10.99	11.66
	Completeness	8.84	12.70	13.57	12.67
	Overall	10.19	11.97	12.28	12.17
Hzau-5	Accuracy	11.17	11.23	10.69	12.32
	Completeness	8.21	13.04	13.04	13.73
	Overall	9.69	12.14	11.86	13.02

Note: The Wuhu-2 and Wuhu-3 use RGB-D point clouds as the ground truth, while Hzau-2 to Hzau-5 use LiDAR data as the ground truth.

4.5. Orchard phenotype measurement experiment

Using high–accuracy LiDAR scans as ground truth, we first performed a quantitative quality assessment of the point clouds generated by four state–of–the–art 3D reconstruction methods, evaluating their precision, completeness, and overall quality. As shown in Table 3, InspectGaussian outperformed all other methods across all metrics, with particularly notable advantages in point cloud completeness. Overall, while maintaining high reconstruction precision, InspectGaussian significantly improves scene completeness and overall reconstruction quality, demonstrating well–rounded performance suitable for 3D phenotyping in orchard environments.

We further evaluated the plant phenotyping measurements using the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and prediction bias. Detailed visualizations of these correlations and error distributions are provided in Fig. S2 of the Supplementary Material. As presented in Fig. S2(A) and (C), plant height prediction achieved a high R² of 0.802, with an RMSE of 9.76 cm and an MAE of 5.76 cm, indicating excellent predictive accuracy. The primary branch angle prediction also attained an R² of 0.778, with acceptable error levels (RMSE = 4.92°, MAE = 4.26°). Moreover, the residual plots (Fig. S2(B) and (D)) show that the prediction errors for both traits are randomly distributed around zero without obvious systematic bias, with mean ± standard deviation values of 0.64 ± 9.74 cm for plant height and 0.82 ± 4.85° for branch angle. These results confirm that the proposed 3D reconstruction and measurement pipeline is reliable for quantifying key plant phenotypic traits and can provide an accurate data foundation for subsequent high–throughput phenotyping analysis.

5. Conclusions

This study introduced InspectGaussian, a coarse-to-fine Gaussian reconstruction framework designed for large-scale orchard inspection. By extending ORB-SLAM3 with dense mapping, we generated real-world–scale point clouds that preserve both geometric structure and semantic detail. A divide-and-conquer strategy was then applied to individual plant reconstruction: YOLO-World–based viewpoint extraction, combined with camera pose estimation, enabled reliable plant-specific image sequences, while an enhanced 3D Gaussian Splatting pipeline with depth regularization and region-aware optimization improved reconstruction fidelity. Experimental validation in citrus orchards demonstrated that InspectGaussian achieves sub-centimeter geometric accuracy, high-quality appearance modeling, and superior robustness compared to existing methods, even under occlusion, illumination variability, and large-scale scene complexity. The results highlight its potential for practical high-throughput phenotyping and smart orchard management.

Nonetheless, reliance on QR code markers introduces additional labor requirements. Future research will focus on eliminating marker dependency by improving natural feature recognition and pose estimation, further enhancing adaptability, automation, and scalability. Ultimately, this work provides a foundation for next-generation intelligent orchard monitoring systems that integrate efficiency, precision, and resilience in unstructured agricultural environments.

Author contributions

Y. Chen, L. Zhang, and Z. Bao contributed to the overall framework and conceptualization. Z. Bao and L. Zhang conducted dataset collection, algorithm development, and experimental implementation. Y. Chen, H. Kang, W. Li, and H. Chen supervised the experiments at all stages and revised the manuscript.

Funding

This work was supported by the National Key Research and Development Program of China under Grant 2023YFF1000100; the Fundamental Research Funds for the Central Universities in China under Grant 2662024SZ002 and Grant 2662024XXPY001.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

The data supporting the findings of this study can be obtained from the corresponding author upon reasonable request. Our code are available at https://github.com/zlhzau/InspectGaussian.git.