Block-Level Index Mixing and Classification Enhancement Attention for occluded person re-identification

Abstract

In the field of occluded person re-identification (Re-ID), a major challenge lies in effectively recognizing partially occluded pedestrian images. Existing methods leveraging occlusion simulation and local feature segmentation often struggle with real-world complexity and inadvertently introduce noise from non-target regions. To address these limitations, this paper proposes the Block-level Index Mixing and Classification Enhancement Attention (BMCE) framework, which integrates data augmentation and classification enhancement strategies. For data augmentation, the Block-level Index Mixing (BLIM) module partitions images with different labels into several blocks. By sharing an index list and controlling the proportion of sampled blocks, the module simulates diverse occluded images. Additionally, adaptive weight mixing of labels enhances the discriminative ability of the simulated images. For classification enhancement, the Classification Enhancement Attention (CEA) module leverages multi-granularity features to enhance classification weights and mitigates the influence of non-target regions through an attention mechanism, improving performance in occlusion scenarios. Experimental results demonstrate that BMCE achieves competitive performance on occlusion, partial, and holistic Re-ID datasets. Notably, it attains 74.1% Rank-1 accuracy and 64.7% mAP on the Occluded-Duke dataset. Source code is available at https://github.com/aohan-del/BMCE.