Local Cross-Patch Activation From Multi-Direction for Weakly Supervised Object Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) aims to identify object locations using only image-level labels.
• Transformers have been applied to WSOL to improve feature dependency but face challenges with background over-activation and occluded object localization.
• Existing methods struggle to differentiate similar object boundaries from backgrounds and capture local features across occlusions.

Purpose of the Study:

• To propose a novel transformer-based WSOL method, LCA-MD, that addresses background over-activation and incomplete activation of occluded objects.
• To enhance the capture of local features while suppressing background noise.
• To improve the accuracy and robustness of object localization in weakly supervised settings.

Main Methods:

• Developed LCA-MD, a transformer-based WSOL approach utilizing local cross-patch activation from multiple directions.
• Introduced a token feature contrast module (TCM) combining contrastive learning with transformers to maximize foreground-background distinctions.
• Proposed a semantic-spatial fusion module (SFM) for capturing local cross-patch features and enabling activation diffusion across occlusions.

Main Results:

• LCA-MD demonstrated significant superiority over existing methods on the CUB-200-2011 and ILSVRC datasets.
• Achieved state-of-the-art performance in weakly supervised object localization.
• The proposed modules effectively inhibited background over-activation and improved localization of occluded objects.

Conclusions:

• LCA-MD offers an effective solution for the challenges in transformer-based WSOL.
• The method enhances the model's ability to capture fine-grained local details and semantic context.
• LCA-MD represents a significant advancement in weakly supervised object localization research.