Automated Ranking of Soybean Plots from Close-Range RGB Video via Depth Filtering and Point-Based Counting

Manual assessment of soybean yield components, such as pod number, is laborious, time-consuming, and subjective. Existing computer-vision approaches based on object detection or instance segmentation perform poorly on close-range RGB imagery of soybean canopies due to severe occlusions, ambiguous plant boundaries, and the high cost of bounding-box annotation. To address these challenges, we propose a video-based ranking pipeline that avoids explicit per-plant detection and instead aggregates global frame-level features derived from monocular depth estimation, semantic plant segmentation, and pod keypoint counting. Depth maps from Depth Anything 3 are used to suppress background clutter, while zero-shot segmentation with SAM3 (distilled into a lightweight SegFormer) provides plant-area signals and a point-based counting network (P2PNet) estimates pod counts. We evaluate the method on a controlled dataset comprising 11 ranks with multiple indoor and outdoor scenes. The proposed pipeline provides a practical, cost-effective solution for automated bed-quality scoring, although it currently depends on fixed camera geometry and does not yet associate pods with individual plants.