WS-RCNN: Learning to Score Proposals for Weakly Supervised Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Weakly supervised instance segmentation (WSIS) addresses the challenge of limited labeled data.
• Existing WSIS methods often rely on heuristic proposal scoring, limiting progress.
• A learning-based approach to proposal scoring is needed for advancing WSIS.

Purpose of the Study:

• To introduce a novel framework, Weakly Supervised R-CNN (WS-RCNN), for weakly supervised instance segmentation.
• To develop a deep network for learning proposal scores under weak supervision.
• To enhance the robustness and accuracy of WSIS.

Main Methods:

• Proposed a novel framework: Weakly Supervised R-CNN (WS-RCNN).
• Introduced Attention-Guided Pseudo Labeling (AGPL) to generate proposal-level pseudo labels using image-level attention maps.
• Developed an Entropic OpenSet Loss to effectively handle background proposals.

Main Results:

• WS-RCNN significantly outperforms state-of-the-art methods on standard datasets.
• Achieved improvements of 11.6% (PASCAL VOC 2012) and 10.7% (MS COCO 2014) in mAP50.
• Demonstrated the effectiveness of learning-based proposal scoring for WSIS.

Conclusions:

• The proposed WS-RCNN framework offers a promising direction for weakly supervised instance segmentation.
• Learning-based proposal scoring is crucial for advancing WSIS.
• AGPL and Entropic OpenSet Loss contribute to improved performance and robustness.