In this paper, we propose an integrated three-dimensional visual perception pipeline that integrates tea bud recognition, semantic segmentation, counting, and harvesting-oriented candidate picking point estimation to support harvesting planning and preparation in real tea plantation environments. The overall workflow of the proposed framework is illustrated in Figure 1.

The selection of YOLOv11, SAM2, Depth Anything V2, and NeRF was guided by a balance between accuracy, robustness, and practical applicability in complex outdoor harvesting environments. YOLOv11 provides strong performance for small-object detection with high efficiency, while SAM2 enables flexible and accurate segmentation with minimal annotation overhead. Depth Anything V2 offers reliable monocular depth estimation with strong cross-scene generalization, which is well-suited for natural environments. NeRF, combined with depth supervision and semantic rendering, is adopted to obtain spatially coherent fine-scale geometry under repetitive textures and self-occlusion, providing a coherent three-dimensional representation that better supports subsequent harvesting-oriented analysis.

This study focuses on tea bud image acquisition in complex outdoor environments, and all data were collected autonomously. To improve acquisition efficiency and satisfy the requirements of subsequent 3D reconstruction, an automated image acquisition system was developed. The device enables controlled rotation around individual tea trees and flexible adjustment of the number of captured images by setting the acquisition interval.

All images were captured using an Obsmeet 4K camera, with a resolution of 3840 × 2160 pixels and stored in JPG format. Data collection was conducted in April under natural field conditions in Chongyang County, Xianning City, Hubei Province, China. The final dataset contains 4,700 static images of yellow tea buds, covering diverse viewpoints and occlusion conditions, as illustrated in Figure 2. Tea bud targets in the images were manually annotated using the Trex annotation tool to generate bounding-box labels for model training.

The dataset used in this study was collected from a single geographic region and tea variety, which limits direct evaluation of cross-region and cross-variety generalization. The primary objective of this work is to validate the feasibility and effectiveness of a harvesting-oriented three-dimensional perception pipeline under controlled field conditions, rather than to establish generalization across different tea cultivars, growth stages, or environmental settings. Systematic evaluation under more diverse regions, tea varieties, and growth conditions will be explored in future work.

In the data preprocessing stage, COLMAP was first employed to estimate the camera poses (including position and orientation) of the original images, thereby restoring the geometric information during image acquisition; in this work, COLMAP is used for camera pose estimation rather than dense surface reconstruction.

Meanwhile, to further enhance data diversity and improve the generalization of the YOLO-based tea bud recognition model under the available training data conditions, various data augmentation techniques were applied to the dataset, as shown in Figure 3. These augmentations were only used during the training of the two-dimensional recognition model.

Specifically, the applied methods included converting images to grayscale and adjusting brightness to approximate different illumination appearances, rather than explicitly modeling real-world illumination variations. Random noise was also synthetically added to increase data variability and expose the recognition model to diverse perturbations during training. In addition, random hue and saturation adjustments were applied to increase color diversity and improve generalization in complex scenes.

In contrast, for the three-dimensional reconstruction process, data augmentation and noise injection were not applied. Instead, images were further screened based on quality indicators such as lighting conditions and sharpness. Low-quality images—such as those affected by overexposure, underexposure, or low contrast—were manually removed to ensure the geometric stability of camera pose estimation and subsequent neural reconstruction, rather than to optimize any specific reconstruction method. After selection, the number of high-quality images retained in each group ranged from 168 to 225.

To reconstruct high-quality three-dimensional representations of tea trees, we adopt the Nerfacto framework within NeRFStudio (Mildenhall et al., 2021; Zhang et al., 2021). Nerfacto provides an efficient implementation of neural radiance fields by optimizing the sampling strategy and network design, achieving a favorable balance between rendering quality and computational efficiency. This makes it suitable for high-resolution reconstruction of complex plant structures under field conditions.

To further enhance geometric fidelity and accelerate convergence, monocular depth priors are incorporated during training through a depth-supervised loss, enabling the model to better capture fine-scale structures such as thin branches and densely clustered tea buds. In addition, a semantic branch is introduced to extend Nerfacto from pure appearance modeling to semantic-aware 3D reconstruction. The semantic field predicts tea bud probabilities as a function of spatial location, allowing semantic information to be consistently aligned with reconstructed geometry. Standard volumetric rendering formulations for color, depth, and semantics are consistent with the original NeRF framework and are therefore not detailed here.

Unlike most existing NeRF-based agricultural studies that primarily focus on visual reconstruction or phenotypic analysis, this section introduces a harvesting-oriented three-dimensional perception strategy. By explicitly processing bud-level semantic point clouds, the proposed method bridges three-dimensional reconstruction with practical harvesting tasks, including tea bud counting and harvesting-oriented candidate point estimation. The harvesting-oriented candidate point is obtained from bud-level semantic point clouds and provides a three-dimensional spatial reference for subsequent harvesting planning. This design enables the extraction of actionable spatial cues from reconstructed point clouds, establishing an effective perception foundation for harvesting-oriented applications.

The detection models in this study were evaluated using precision (P), recall (R), and mean accuracy (mAP), where P denotes the proportion of accurate predictions in all prediction examples and R denotes the proportion of accurate predictions in all true examples. mAP denotes the composite accuracy metric used to evaluate the detection models. The formulas for the computation of P, R, and mAP are shown in Equations (6–8).

TP: number of positive samples predicted as positive samples.

FP: Number of negative samples predicted as positive samples.

FN: number of positive samples predicted as negative samples.

N: denotes the number of bud types detected (only one type of tea bud is studied in this paper, so N is equal to 1).

It can be seen from Table 1 that each component, including DySample, DG-SimAM, and InnerIoU, contributes positively to the overall detection performance. Specifically, DySample mainly improves detection coverage, as reflected by the increase in mAP@50, while InnerIoU enhances localization accuracy, resulting in a higher mAP@50:95. DG-SimAM further improves recall and robustness under complex backgrounds.

When combined, the proposed model achieves the best results, with Precision, Recall, mAP@50, and mAP@50:95 reaching 0.827, 0.843, 0.917, and 0.651, respectively. Compared to the baseline YOLOv11n, these values represent improvements of 5.1% in Precision, 4.2% in Recall, 4.8% in mAP@50, and 2.7% in mAP@50:95, while maintaining comparable parameters and computational cost (GFLOPs). This indicates that the introduced modules not only enhance feature extraction and localization accuracy, but also achieve a favorable trade-off between accuracy and efficiency.

Furthermore, as summarized in Table 2 and illustrated in Figure 11, the proposed model consistently outperforms other YOLO variants across all evaluation metrics. Although YOLOv5n, YOLOv8n, and YOLOv10n demonstrate competitive performance, the improved YOLOv11-based model achieves superior detection accuracy and localization stability. In particular, the improvement in mAP@50:95 indicates more precise bounding box regression under stricter IoU thresholds, which is crucial for detecting small, densely distributed, and partially occluded tea buds in complex tea garden environments.

Accurate localization is especially important in the proposed pipeline, as the detection results are directly used as prompts to guide SAM2 for fine-grained semantic segmentation. Inaccurate or shifted bounding boxes may propagate errors to the segmentation stage, leading to incomplete or imprecise tea bud masks.

To achieve semantic segmentation of tea buds, we employed three methods: YOLO+SAM, YOLO+SAM2, and a self-trained U-Net (Ronneberger et al., 2015). In addition, a representative single-stage instance segmentation model, Mask R-CNN, was included as a baseline for comparison. The U-Net training data set consisted of two parts: 60 manually annotated images from our captured data set and 30 additional images refined from the YOLO+SAM2 segmentation results. Data augmentation techniques, including color transformation (brightness, contrast, and saturation adjustments within 0.5–1.5), image flipping, rotation, scaling, and noise injection, were applied to improve robustness. The U-Net model was implemented in PyTorch.

The comparative performance of all methods is summarized in Table 3. YOLO+SAM2 achieved the highest IoU (0.640) and Dice coefficient (0.779) while maintaining a reasonable inference time (0.511 s). YOLO+SAM closely followed with an IoU of 0.629 and a Dice score of 0.771. U-Net was significantly faster (0.013 s) but less accurate (IoU 0.597, Dice 0.747). Mask R-CNN exhibited lower segmentation accuracy (IoU 0.578, Dice 0.725) while remaining computationally efficient (0.028 s).

As shown in Figure 12, YOLO+SAM2 produces more precise segmentation of small and occluded tea buds, with improved boundary delineation compared to SAM, benefiting from its more efficient attention mechanism and lighter architecture. YOLO+SAM also shows competitive performance, though with slightly reduced accuracy. U-Net, despite its high speed and independence from cue information, tends to over-segment tea leaves and struggles with fine or deeply embedded tea buds under complex canopy conditions. Mask R-CNN, while capable of directly predicting instance-level masks, often suffers from incomplete or fragmented bud segmentation in densely occluded regions, which limits its effectiveness in this scenario.

Overall, the quantitative results in Table 3 and the qualitative comparisons in Figure 12 demonstrate that YOLO+SAM2 achieves the best balance between segmentation accuracy and inference efficiency. Although segmentation errors remain under realistic harvesting conditions—mainly due to boundary ambiguity and severe occlusion—the improved boundary precision of YOLO+SAM2 leads to more compact and consistent semantic point clusters, which is beneficial for downstream tasks such as tea bud counting and candidate picking-point estimation in harvesting-oriented 3D perception.

Semantic 3D reconstruction is a key part of the process. The classical 3D representations of point clouds, meshes, and voxels are limited in their effectiveness in representing complex scenes and are limited in representing fine-grained geometry and view-dependent appearance in densely cluttered canopies, motivating the use of NeRF to model radiance and density as continuous fields. With body rendering techniques, NeRF can generate high-quality images from arbitrary viewpoints. Nerfstudio is an open-source framework that provides many frameworks for NeRF. We extend the Nerfacto method by adding a semantic layer that maps points and viewpoint directions in 3D space to semantic information as well, which can be viewed as a semantic field.

As shown in Figure 13, the proposed 3D reconstruction framework for tea trees adopts multimodal encoding and collaborative neural network modeling to establish an efficient mapping between 3D space and 2D images at the reconstruction level. Specifically, spatial points (x, y, z) along each ray and their corresponding view direction vectors d are encoded using hash encoding and spherical harmonic (SH) encoding for feature extraction, respectively. Hash encoding effectively reduces the computational overhead of traditional positional encoding and supports high-resolution geometric modeling, while SH encoding captures view-dependent orientation information to improve the modeling accuracy of lighting and material appearance. The fused appearance embedding vectors are nonlinearly mapped by a multilayer perceptron (MLP) to predict volume density (σ) and color (RGB), and novel views are rendered via volume rendering. This decoupled architecture improves reconstruction efficiency compared to conventional NeRF formulations while maintaining reconstruction quality. On an NVIDIA RTX 4090 (24GB VRAM), training on a set of 3840×2160 resolution images takes approximately 5 minutes per tea tree. This runtime reflects an offline, per-tree reconstruction setting focused on high-fidelity geometric and semantic perception, rather than real-time harvesting execution.

For completeness, an approximate runtime breakdown of the proposed pipeline is provided in the Supplementary Material (Supplementary Table 2).

In the training phase, the photometric loss is calculated as defined in Equation (9) by comparing the difference between the predicted color of each ray and the real pixel RGB value, thus optimizing the color reconstruction performance of the model. For semantic rendering, the semantic loss is defined in Equation (10), where the model estimates the probability of each pixel belonging to the tea bud or the background based on the prediction density and introduces the binary cross-entropy loss to enhance semantic segmentation accuracy. Meanwhile, to enhance the model’s ability to model the geometric structure of the scene, the depth-supervised loss is defined in Equation (11), improving the accuracy and robustness of 3D reconstruction by constraining the consistency between the predicted depth and the true depth.

The overall training objective is defined in Equation (12):

In this study, NeRF is applied to achieve 3D reconstruction of tea trees based on multi-view images. The model synthesizes geometrically consistent views from different perspectives, thereby generating dense point clouds that support subsequent structural and semantic analysis. Nevertheless, conventional NeRF often produces rendered images that appear less sharp than the original photos, particularly in regions with densely distributed tea buds. This is mainly because NeRF prioritizes geometric consistency and multi-view synthesis fidelity, while fine-grained details may accumulate rendering errors and lead to local blurring. To overcome this limitation, depth information was incorporated into NeRF (Depth-NeRF), allowing more accurate depiction of subtle tea bud structures, clearer leaf-edge contours, and improved overall sharpness, as illustrated in Figure 14.

To systematically evaluate the contribution of monocular depth priors to Neural Radiance Fields (NeRF) for reconstructing plant-like natural objects, three tea trees with distinct morphologies (denoted as Tree1, Tree2, and Tree3) were selected. The baseline NeRFacto model was compared with a depth-supervised variant, Depth-NeRF, which incorporates monocular depth priors estimated using Depth Anything v2 during training iterations ranging from 2,000 to 30,000. Reconstruction performance was evaluated using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS), providing a comprehensive assessment of both image fidelity and perceptual quality.

As shown in Table 4, Depth-NeRF consistently outperforms NeRFacto in all three test cases. In the Tree1 scene, Depth-NeRF achieves a PSNR of 24.08, SSIM of 0.704, and LPIPS of 0.209 at 30,000 iterations, compared to 23.91, 0.698, and 0.247 for NeRFacto, respectively. Although the absolute PSNR and SSIM improvements appear modest, the substantial reduction in LPIPS highlights a clear enhancement in perceptual realism and fine structural detail. A similar trend is observed for Tree2 and Tree3, with Tree3 showing the most significant gain (LPIPS reduced from 0.599 to 0.405), indicating improved robustness in handling occlusion, texture repetition, and complex leaf structures. However, residual reconstruction uncertainty remains in densely occluded regions and around thin structures, where monocular depth ambiguity may cause local geometric blur or incomplete surfaces that can propagate to subsequent semantic point cloud extraction.

Another important observation is that Depth-NeRF exhibits faster convergence during the early training phase (2k–8k iterations), producing sharper and less noisy reconstructions compared to the baseline. This demonstrates that monocular depth priors not only improve final reconstruction quality but also stabilize training and accelerate geometry learning under limited iterations.

Figure 15 further illustrates the qualitative improvements. The RGB renderings generated by Depth-NeRF under natural lighting accurately restore the tea tree morphology, with clearer branch-leaf hierarchies than NeRFacto. In semantic rendering, tea buds (highlighted in red) are accurately separated from surrounding leaves (blue-green), verifying the capability of Depth-NeRF to achieve fine-grained semantic segmentation in cluttered environments. The depth maps reveal an enhanced contrast between the body of the tea tree (blue-to-yellow gradient) and the background (orange-red), strengthening the model’s ability to capture meaningful depth differences.

Overall, the integration of monocular depth priors significantly improves both perceptual realism and structural fidelity in the reconstruction of NeRF-based plants. These improvements are particularly valuable for downstream applications, such as tea bud counting and candidate picking-point estimation for harvesting- oriented perception, where precise structural representation and semantic separation are essential.

After training the neural radiation field model, we utilize the volume sampling module provided by Nerfstudio to reconstruct the scene in three dimensions and generate high-quality point cloud data that integrate geometric structure, appearance color, and semantic information. The model is designed with three cooperative components. The Density Field encodes the volume density of each spatial point, which enables effective separation of solid surfaces such as tea tree leaves and branches from the background. The Appearance Field associates each point with RGB color information to recover realistic lighting and material properties. The Semantic Field predicts the semantic probabilities of key targets such as tea buds, thereby supporting semantic-level 3D reconstruction.

Nevertheless, during neural field training, semantic features may propagate along the line-of-sight direction, and direct sampling of the semantic field often introduces noisy labels in invalid regions, such as empty background space. To overcome this limitation, we propose a density-symmetric Coupling (DSC) strategy. In this method, a density threshold (σ ≥ 0.45) is applied and semantic predictions are only retained for spatial points where the density field indicates the presence of solid surfaces. This constraint effectively suppresses spurious semantic predictions in invalid space and ensures that semantic labels remain consistent with actual physical structures, as shown in Figure 16. By constraining semantic predictions with density responses, DSC suppresses a major source of structured semantic noise and improves the reliability of downstream clustering and counting.

Based on the reconstructed semantic point cloud, tea bud counting and harvesting-oriented candidate point estimation are performed as perception-level outputs to support downstream harvesting tasks. Once the point cloud data of the tea buds are obtained, a clustering algorithm is applied to structure the data for tea bud counting and candidate point estimation. Before clustering, the point clouds are preprocessed using Statistical Outlier Removal and Radius Outlier Removal to eliminate noise and anomalies, thereby improving the reliability of subsequent analysis. In practice, multiple small tea buds located in proximity may be geometrically merged into the same cluster, leading to biased counting results. To alleviate this issue, a DBSCAN-based clustering strategy with an adaptive neighborhood radius (guided by local point density) and a normal-consistency constraint (clamp angle ≤ 15^°) is adopted, which helps reduce cluster adhesion in dense bud regions.

Following clustering, the RANSAC algorithm is applied to fit the dominant growth plane of each tea bud cluster. A distance threshold is then used to retain in-plane points, which are subsequently projected along the normal vector of the fitted plane. The point with the smallest projection value is selected as the harvesting-oriented candidate point for the corresponding tea bud cluster. It should be noted that this candidate point represents a perception-level three-dimensional guidance cue derived from geometric structure, rather than an anatomically exact stem–bud junction or a robot-executable cutting command.

To illustrate the feasibility of three-dimensional tea bud counting based on semantic point clouds under different occlusion conditions, three individual tea trees are selected as representative case studies. These trees are not randomly sampled; instead, they are deliberately chosen to reflect increasing levels of occlusion complexity, ranging from relatively sparse foliage to severe self-occlusion. The tea buds on each tree are manually harvested, and the corresponding counts are recorded as reference values. Three-dimensional reconstruction is then performed using NeRF, and the resulting semantic point clouds are used for clustering-based counting. The qualitative correspondence between reconstructed bud clusters and manually observed buds is visualized in Figure 17, while the case-level counting results are summarized in Table 5. This experiment is intended as an illustrative, case-level evaluation rather than a population-level statistical analysis.

To quantitatively assess the localization reliability of candidate picking points under realistic conditions, we employ a proxy-based evaluation strategy that focuses on geometric stability under repeated perturbations. Specifically, for each tea bud cluster, the candidate picking point is re-estimated multiple times under random subsampling of the point cloud, and the variability of the resulting points is measured.

It should be clarified that this proxy-based evaluation does not aim to directly measure biological correctness with respect to manually annotated stem–bud junctions. Due to the lack of reliable junction-level ground truth for real tea plants at scale, the adopted evaluation focuses on repeatability and consistency, which are necessary properties for harvesting-oriented guidance cues used in downstream planning.

Three quantitative metrics are adopted to assess localization stability: (1) the mean point deviation µ_p, which measures the average Euclidean distance between repeated estimations; (2) the 90th percentile deviation p90_p, which reflects worst-case sensitivity under perturbation; and (3) the pass rate, defined as the proportion of estimations whose deviation falls below a predefined tolerance threshold. Lower values of µ_p and p90_p, together with a higher pass rate, indicate more stable and reliable candidate point estimation.

Table 6 summarizes the proxy-based localization stability results across three individual tea trees under consistent experimental settings (p = 5, subsample ratio = 0.5, noise = 0). The proposed method (ours) achieves substantially lower spatial deviation than PCA-based baselines and demonstrates competitive stability compared with centroid-based approaches. Although the cluster centroid method yields the lowest numerical deviation and the highest pass rate, it tends to favor geometrically central locations that lack explicit harvesting-oriented structural cues, as it does not incorporate any basal-side or attachment-related priors of tea buds.

To further analyze the sensitivity of the proposed method to candidate point displacement, an offset strategy is introduced, where the estimated candidate picking point is shifted along the negative normal direction of the fitted local plane by a scaled distance Δk. As Δk increases from 1.5 to 5.0, both µ_p and p90_p increase monotonically, accompanied by a gradual decrease in pass rate. This trend indicates that excessive displacement amplifies instability under perturbations, highlighting the importance of moderate offsets when incorporating harvesting-oriented geometric priors.

Overall, the results demonstrate that the proposed method provides a stable and task-consistent harvesting-oriented guidance cue under geometric perturbations. While centroid-based approaches achieve strong numerical stability, the proposed strategy integrates basal-side geometric priors that are more aligned with harvesting-oriented localization. Its localization objective focuses on harvesting-oriented geometric referencing rather than precise anatomical reconstruction of the stem–bud junction. The offset analysis further clarifies the trade-off between robustness and harvesting-oriented displacement in candidate point estimation.

The experimental results demonstrate that the proposed tea bud segmentation framework benefits from the complementary strengths of the large-scale foundation model SAM2 and the lightweight detector YOLOv11. Tea bud segmentation remains a challenging task due to limited annotated data and severe occlusion among buds and surrounding leaves, which often leads to boundary ambiguity and missed detections in complex canopy environments.

While SAM2 exhibits strong generalization capability on unseen samples, its performance on small-scale targets such as tea buds is constrained in the absence of effective localization cues. Conversely, the YOLOv11-based detector provides reliable coarse localization but may suffer from incomplete boundaries under heavy occlusion. By introducing YOLOv11 detection results as prompts for SAM2, the proposed fusion strategy reduces segmentation ambiguity for small and partially occluded buds, leading to improved segmentation consistency and reduced annotation effort.

Nevertheless, segmentation errors are not entirely eliminated. Residual over-segmentation and under-segmentation persist in regions with extreme occlusion or weak visual contrast, and such errors may propagate to subsequent 3D reconstruction stages. In particular, missed or truncated bud regions in 2D segmentation cannot be recovered in later processing, highlighting the critical role of segmentation recall in harvesting-oriented 3D perception pipelines.

Most existing tea bud studies focus on recognition and localization in two-dimensional images, where occlusion and perspective effects fundamentally limit reliable structural analysis. By introducing a NeRF-based reconstruction framework with monocular depth priors, this study extends tea bud perception into three-dimensional space, enabling bud counting and candidate picking-point estimation from a geometric perspective.

Incorporating depth information derived from Depth Anything v2 into the Nerfacto framework improves reconstruction sharpness and structural fidelity, particularly during the early stages of training and in regions with complex leaf–bud interactions. From a methodological perspective, NeRF-based reconstruction provides a more flexible representation for modeling fine-scale plant organs with severe self-occlusion and repetitive textures, which often pose challenges to traditional structure-from-motion pipelines (e.g., COLMAP) by degrading feature matching and geometric consistency.

Nevertheless, reconstruction uncertainty remains an inherent challenge when modeling tea trees with fine-grained geometry and heavy self-occlusion. Monocular depth estimation is prone to ambiguity around thin structures and densely layered foliage, leading to local geometric blur or incomplete surfaces in the reconstructed space. Such errors are spatially correlated rather than random and may propagate into the semantic field and subsequent point cloud extraction, although depth supervision substantially suppresses reconstruction noise.

While monocular depth supervision improves the overall geometric consistency of NeRF reconstruction, it is inherently subject to estimation errors in visually ambiguous regions such as thin branches and dense foliage. To quantitatively assess how such depth errors may affect the reconstructed geometry, we conduct a region-aware depth reliability analysis by comparing Depth Anything V2 predictions with sparse depth estimates derived from COLMAP reconstruction.

Specifically, sparse 3D points reconstructed by COLMAP are projected onto the image plane, and their camera-space depths are treated as reference depth samples. At corresponding pixel locations, monocular depth predictions are extracted and aligned using robust median scaling to account for the inherent scale ambiguity of monocular depth estimation. Depth errors are then evaluated for all valid samples (Global), manually selected ambiguous regions of interest (ROIs), including thin branches and dense foliage, and non-ROI regions as a control group.

Quantitative results are summarized in Table 7. The global region exhibits moderate depth errors, with performance comparable to that of non-ROI regions, indicating that monocular depth supervision provides reasonably consistent structural guidance at the scene level. In contrast, error magnitudes increase noticeably in visually ambiguous regions. Thin-branch ROIs show moderately higher errors, reflecting the difficulty of estimating depth for fine, elongated structures. Dense-foliage ROIs exhibit substantially larger absolute and relative errors, together with significantly heavier-tailed error distributions, as evidenced by the elevated p90 AbsRel values. This behavior can be attributed to severe occlusion, multi-layer leaf overlap, and weak monocular depth cues in dense foliage areas.

These results indicate that depth estimation errors are not uniformly distributed across the scene, but are spatially concentrated in regions with inherent visual ambiguity. Consequently, monocular depth supervision may introduce localized geometric blur or noise in such regions, while preserving global structural fidelity. This region-dependent error characteristic explains the localized reconstruction artifacts observed in dense foliage areas and supports the subsequent stability analysis of candidate picking point estimation.

Most existing tea bud yield estimation approaches rely on two-dimensional images, where occlusion and projection overlap frequently lead to counting ambiguity. By leveraging a reconstructed semantic point cloud, the proposed method enables tea bud counting from a three-dimensional perspective, which helps alleviate ambiguity caused by overlapping buds in 2D views. Rather than serving as a population-level statistical evaluation, the counting results presented in this study are intended as a case-level illustration to examine the feasibility and behavior of 3D-based counting under different occlusion conditions.

As observed in the representative cases, discrepancies between reconstructed counts and manual reference values still occur. Under-counting typically arises when severely occluded tea buds are incompletely reconstructed and fail to form sufficiently dense point cloud clusters, or when closely adjacent buds are geometrically merged into a single cluster. In contrast, partial fragmentation of an individual bud may occasionally lead to over-counting. These error patterns are not randomly distributed; instead, they reflect the stage-wise accumulation of uncertainty across segmentation, depth estimation, and point cloud clustering, rather than isolated inaccuracies at a single processing step.

For harvesting-oriented candidate point estimation, existing approaches predominantly rely on two-dimensional localization cues or point clouds acquired through LiDAR scanning. In contrast, this study derives candidate points directly from reconstructed semantic point clouds by combining bud-level clustering with RANSAC-based local geometric modeling. The resulting candidate points represent perception-level three-dimensional guidance cues that capture the relative spatial structure of tea buds, rather than anatomically exact stem–bud junctions. Although their spatial consistency depends on the completeness of bud reconstruction and the stability of cluster-level geometric fitting, these candidate points provide actionable spatial references for downstream harvesting planning. Systematic validation across larger-scale datasets and explicit occlusion-level stratification are left for future work.