Cross-Domain Person Re-Identification Based on Multi-Branch Pose-Guided Occlusion Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain person re-identification faces challenges with occlusions and fixed model parameters, hindering accurate feature matching.
• Pedestrian feature misalignment due to occlusion is a major obstacle in person re-identification systems.
• Existing methods struggle with generalization and robustness when dealing with occluded individuals in diverse datasets.

Purpose of the Study:

• To propose a novel multi-branch pose-guided occlusion generation method for cross-domain person re-identification.
• To enhance the model's ability to extract discriminative features from non-occluded areas and handle occluded samples effectively.
• To improve identity matching accuracy and generalization performance in challenging re-identification scenarios.

Main Methods:

• A pose-guided occlusion generation module was designed to simulate occluded person images, improving learning on non-occluded features.
• A multi-branch feature fusion structure was employed to enrich feature diversity by combining global and occlusion-specific features.
• Dynamic convolution kernels were utilized for similarity calculation, enabling effective point-to-point matching and overcoming fixed parameter limitations.

Main Results:

• The proposed method demonstrated significant advantages over mainstream algorithms in Rank-1, mean average precision (mAP), and generalization.
• On the MSMT17→DukeMTMC-reID dataset, mAP and Rank-1 reached 80.5% and 84.3% respectively, with Rerank and Tlift achieving 81.9% and 93.1%.
• The algorithm achieved 51.6% and 41.3% on DukeMTMC-reID→Occluded-Duke, showcasing strong recognition performance on occluded datasets.

Conclusions:

• The multi-branch pose-guided occlusion generation method effectively addresses feature matching issues caused by occlusion in person re-identification.
• The approach enhances model generalization and robustness, enabling accurate identity matching even with significant occlusions and feature misalignment.
• This method offers a promising solution for improving person re-identification performance in real-world, occluded scenarios.