Improved Transformer for Person Re-identification with Multi-scale and Multi-granularity Feature Fusion

Person re-identification (ReID) aims to match images of the same individual captured from different cameras or perspectives, which is often hindered by challenges such as viewpoint changes, occlusion, and background clutter. Existing ReID approaches that focus on local feature extraction frequently struggle to capture sufficiently discriminative representations, leading to suboptimal performance. To address these issues, our approach integrates a refined Transformer with feature fusion across various scales and granularities. Specifically, we introduce a local multi-granularity branch designed to capture semantic cues at varying levels of detail, enhancing feature representation by incorporating finer-grained pedestrian attributes. The enhanced Transformer component is capable of concentrating on key regions of interest while simultaneously incorporating global contextual information. To further refine feature learning, we employ a combination of triplet and softmax losses across different branches. Results from experiments on CUHK03, Market1501, and DukeMTMC-reID demonstrate that our approach surpasses several recent leading methods with respect to accuracy.