Looking Beyond Single Images for Weakly Supervised Semantic Segmentation Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation (WSSS) typically relies on image-level labels.
• Existing methods struggle to capture complete object content using intra-image information alone.
• There is a need for methods that leverage cross-image relationships for improved object pattern mining.

Purpose of the Study:

• To develop a novel framework for WSSS that effectively utilizes cross-image semantic relations.
• To enhance the mining of comprehensive object patterns by considering similarities and differences across images.
• To improve the accuracy and completeness of object localization maps for better semantic segmentation.

Main Methods:

• Incorporation of two neural co-attention mechanisms into a classifier.
• Utilizing co-attention to capture common semantics between objects in different images.
• Employing contrastive co-attention to identify unique semantics of unshared objects across images.

Main Results:

• The proposed method significantly improves object pattern learning and localization map inference.
• It establishes new state-of-the-art results across various WSSS settings, including precise, single-label, and noisy web data.
• The approach achieved 1st place in the Weakly-Supervised Semantic Segmentation Track at CVPR2020.

Conclusions:

• Cross-image semantic relations are valuable for comprehensive object pattern mining in WSSS.
• The proposed unified framework effectively handles diverse WSSS scenarios.
• The method demonstrates high utility and efficacy, outperforming existing approaches.