Abstract
Occlusion presents a significant challenge in grape phenotyping detection, where predicting occluded content (amodal detection) can greatly enhance detection accuracy. Recognizing that amodal detection performance is heavily influenced by the segmentation quality between occluder and occluded grape instances, we propose a grape instance segmentation model designed to precisely predict error-prone regions caused by mask size transformations during segmentation, with a particular focus on overlapping regions. To address the limitations of current occlusion synthesis methods in amodal detection, a novel overlapping cover strategy is introduced to replace the existing random cover strategy. This approach ensures that synthetic grape instances better align with real-world occlusion scenarios. Quantitative comparison experiments conducted on the grape amodal detection dataset demonstrate that the proposed grape instance segmentation model achieves superior amodal detection performance, with an IoU score of 0.7931. Additionally, the proposed overlapping cover strategy significantly outperforms the random cover strategy in amodal detection performance.