Spatial Structure Constraints for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation relies on easily available image-level labels.
• Class activation maps (CAMs) offer object location clues but only highlight discriminative parts.
• Existing expansion strategies for CAMs often lead to over-activation in background regions.

Purpose of the Study:

• To propose spatial structure constraints (SSC) to mitigate over-activation in attention expansion for semantic segmentation.
• To enhance the accuracy of object localization in weakly supervised semantic segmentation tasks.
• To develop a method that refines object localization without relying on external saliency models.

Main Methods:

• Proposed a CAM-driven reconstruction module to constrain attention diffusion by preserving image spatial structure.
• Introduced an activation self-modulation module to refine CAMs using regional consistency for finer details.
• Developed a novel approach for weakly supervised semantic segmentation using spatial structure constraints.

Main Results:

• Achieved 72.7% mIoU on PASCAL VOC 2012.
• Achieved 47.0% mIoU on COCO dataset.
• Demonstrated superior performance without external saliency models.

Conclusions:

• Spatial structure constraints effectively alleviate over-activation in attention expansion.
• The proposed method enhances object localization accuracy in weakly supervised semantic segmentation.
• The approach offers a robust solution for semantic segmentation using readily available image-level labels.