Category-Aware Spatial Constraint for Weakly Supervised Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object detection (WSOD) is a growing research area.
• Current WSOD methods often use category-independent region proposals and image-level labels, leading to imprecise boundary localization.
• Existing schemes struggle with accurate object boundary detection due to limitations in score aggregation.

Purpose of the Study:

• To enhance the precision of object boundary localization in weakly supervised object detection.
• To propose a novel Category-aware Spatial Constraint (CSC) scheme for improving WSOD performance.
• To address the limitations of existing methods in aggregating region proposal scores.

Main Methods:

• Developed a Category-aware Spatial Constraint (CSC) integrated into an end-to-end learning framework.
• Incorporated global object shape information using unsupervised Category-specific Pixel Gradient (CPG) maps.
• Introduced multi-center regularization to penalize deviations between cluster centers and high-score proposals.

Main Results:

• The proposed CSC scheme significantly improves weakly supervised object detection performance.
• The method achieves state-of-the-art results on benchmark datasets like Pascal VOC and COCO.
• The approach enhances boundary localization without introducing new learnable parameters or altering CNN structures.

Conclusions:

• The Category-aware Spatial Constraint (CSC) effectively improves weakly supervised object detection.
• Utilizing Category-specific Pixel Gradient (CPG) maps and multi-center regularization enhances object boundary precision.
• The proposed method offers a parameter-free and structure-agnostic improvement for existing WSOD models.