Content-Adaptive Auto-Occlusion Network for Occluded Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification (ReID) is challenging due to partial visibility in images.
• Existing methods often use suboptimal auxiliary models or part-matching strategies.
• Previous occlusion augmentation (OA) methods had fixed policies and random occlusion application.

Purpose of the Study:

• To develop a novel Content-Adaptive Auto-Occlusion Network (CAAO) for improved occluded ReID.
• To dynamically adjust occlusion augmentation based on image content and network training status.
• To overcome limitations of fixed and random occlusion policies in prior OA-based ReID methods.

Main Methods:

• Proposed the Content-Adaptive Auto-Occlusion Network (CAAO) comprising a ReID network and an Auto-Occlusion Controller (AOC).
• AOC dynamically generates optimal occlusion policies based on ReID network feature maps.
• Utilized an on-policy reinforcement learning-based alternating training paradigm to update both ReID and AOC modules.

Main Results:

• Demonstrated the superiority of CAAO on occluded and holistic person ReID benchmarks.
• The proposed method effectively addresses the limitations of fixed and random occlusion augmentation.
• Achieved improved performance in person re-identification tasks under severe occlusion.

Conclusions:

• CAAO offers a significant advancement in occluded person re-identification by intelligently adapting occlusion augmentation.
• The dynamic and content-aware approach leads to more robust and accurate ReID models.
• This work provides a more effective strategy for training ReID networks in occluded environments.