Hierarchical fusion and local-aware transformer for occluded person re-identification.

Abstract

Occluded person re-identification (ReID) is intended to address the problem of matching pedestrians when images of individuals are partially occluded. Recently, Transformer-based methods for occluded person ReID have received considerable attention. However, although existing methods have achieved promising results, most of them do not fully consider the varying contributions of different patches to identity recognition, nor do they sufficiently emphasize the identity information in critical regions. Furthermore, those methods often lack sufficient capability to extract fine-grained local features, making it difficult to fully explore the identity information embedded in various body parts. To resolve the mentioned problems, we propose a Hierarchical Fusion and Local-aware Transformer (HFLAT) for occluded person ReID. Specifically, we firstly design a feature hierarchical fusion module that hierarchizes and fuses the patch feature vector sequence according to the relative importance of each patch to the global feature vector, thereby reinforcing the identity discriminative features of key regions. We then design a feature separation module to distinguish foreground features from background features by employing patch-level saliency analysis, thereby mitigating the negative impact of backgrounds and occlusions on the performance of the model. In addition, we design a local feature extraction module which restricts the range of interactions between the features using a local-aware multi-head attention mechanism, increasing the model's ability to obtain fine-grained local features. Experimental results on the Occluded-DukeMTMC, Occluded-ReID, Market1501, and DukeMTMC-ReID datasets demonstrate that HFLAT reaches the current state-of-the-art performance for occluded person ReID. Specifically, on the Occluded-DukeMTMC and Occluded-ReID datasets, our method achieves the Rank-1 accuracy of 79.6% and 89.8%, respectively, and the mAP of 64.7% and 84.9%, respectively. On the Market1501 and DukeMTMC-ReID datasets, our method achieves the Rank-1 accuracy of 95.9% and 90.6%, respectively, and the mAP of 90.8% and 82.2%, respectively.