Incoherent Region-Aware Occlusion Instance Synthesis for Grape Amodal Detection

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Occlusion significantly hinders accurate grape phenotyping and detection.
• Amodal detection, predicting occluded content, is crucial for improving accuracy.
• Segmentation quality between occluder and occluded grape instances impacts amodal detection.

Purpose of the Study:

• To develop a grape instance segmentation model for precise prediction of error-prone regions in occluded scenarios.
• To introduce a novel overlapping cover strategy for more realistic occlusion synthesis in amodal detection.
• To enhance the accuracy of grape phenotyping through improved amodal detection.

Main Methods:

• Proposed a grape instance segmentation model focusing on overlapping regions and mask transformations.
• Introduced a novel overlapping cover strategy to replace random cover strategies for occlusion synthesis.
• Conducted quantitative comparison experiments on a grape amodal detection dataset.

Main Results:

• The proposed grape instance segmentation model achieved superior amodal detection performance with an IoU score of 0.7931.
• The novel overlapping cover strategy significantly outperformed the random cover strategy in amodal detection.
• The new methods demonstrated enhanced accuracy in predicting occluded grape content.

Conclusions:

• The developed grape instance segmentation model effectively addresses occlusion challenges in phenotyping.
• The overlapping cover strategy provides a more realistic approach to simulating occlusions for amodal detection.
• These advancements significantly improve grape phenotyping accuracy through enhanced amodal detection.