ACFM: Adaptive Channel Feature Matching for Pedestrian Re-Identification

Image misalignment is a significant challenge in the field of pedestrian re-identification. Previous studies typically align pedestrian features using additional models or by leveraging auxiliary information. However, these methods are data-dependent and can fail when dealing with substantial scene variations. To address these challenges, this paper proposes an efficient adaptive channel feature matching (ACFM) strategy. This method adaptively aligns channel feature maps according to image content, enables accurate matching of local features across images, and mitigates the need for supplementary data. Building on the ACFM approach, this paper develops a multi-branch adaptive matching network model. The model incorporates a channel attention mechanism to enhance the representation capacity of individual channels and integrates the ACFM algorithm within local branches to optimize the distance metric computation, effectively capturing and aligning local details. The multi-branch structure is designed to handle both global and local features, enabling the network to comprehensively capture and integrate image information. Evaluation experiments conducted on two widely-used benchmark datasets, Market-1501 and DukeMTMC-reID, demonstrate that the proposed method offers significant advantages over current state-of-the-art approaches.