Incremental Generative Occlusion Adversarial Suppression Network for Person ReID.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-id) is challenged by occlusions, reducing pedestrian distinctiveness.
• Existing methods often use complex modules or limited data augmentation, yielding suboptimal performance.

Purpose of the Study:

• To develop a robust person re-identification method that effectively handles occlusions.
• To improve feature learning by focusing on non-occluded body regions and spatial misalignment.

Main Methods:

• Proposed the Incremental Generative Occlusion Adversarial Suppression (IGOAS) network.
• Incorporated an incremental generative occlusion block for gradual learning of occlusion patterns.
• Implemented a global-adversarial suppression framework with global and adversarial suppression branches to extract robust features.

Main Results:

• The IGOAS network demonstrated competitive performance on occluded datasets.
• Achieved 60.1% Rank-1 accuracy and 49.4% mAP on the Occluded-DukeMTMC dataset.
• The proposed method shows significant robustness to occlusion problems in person re-identification.

Conclusions:

• The IGOAS network effectively addresses the challenge of occlusion in person re-identification.
• The combination of incremental occlusion generation and adversarial suppression leads to more discriminative and robust pedestrian feature descriptors.