Weakly Supervised Instance Segmentation

Weakly supervised instance segmentation aims to segment individual objects in images using minimal annotation, typically bounding boxes or image-level labels, rather than expensive pixel-level masks. Current research focuses on improving the quality of automatically generated pseudo-masks through techniques like incorporating depth information, refining proposals with mask IoU scores, and leveraging multimodal data (e.g., LiDAR, optical flow) or text recognition outputs. This field is significant because it drastically reduces the annotation burden for training instance segmentation models, enabling broader application in areas like autonomous driving and medical image analysis where large, fully annotated datasets are scarce.