Decouple Re-identification and Human Parsing for Occluded Person Re-Identification

Shuguang Dou, Yuanpeng Tu, Yanna Lin, Junyao Gao, Cairong Zhao

Abstract: This paper proposes a Decouple Re-identificatiOn and human Parsing (DROP) method to learn the task-specific features that fit the two tasks for occluded person re-identification (ReID). Previous multi-task learning approaches for occluded ReID typically either (1) jointly learn person ReID and human parsing from global features, or (2) leverage semantic information from parsing to guide attention, with the latter generally achieving superior performance. The paper attributes this discrepancy to the differing granularity requirements of ReID and human parsing features. To address this, we propose DROP, a framework that explicitly decouples ReID and parsing features. Specifically, we firstly introduce a Hierarchical Bias-Decoupled Adapter (HBDA) to extract task-specific representations for human parsing. To further separate the two tasks, human positional information is incorporated only into the parsing branch with a pedestrian position encoder to provide semantic spatial context, while the ReID branch employs a Part-aware Compactness Triplet (PCT) loss to enhance part-level discriminability at the instance level. Experimental results underscore the efficacy of DROP compared to the two prevailing mainstream methods, especially the Rank-1, which reached 77.2% on Occluded-Duke. The codebase of DROP will be available.

External IDs:doi:10.1109/tbiom.2026.3680760