Occluded Person Re-Identification Model using Swin Transformer with Multi-Scale Features

Person re-identification is one of the important research directions in computer vision, but existing methods face issues such as low recognition accuracy in occluded scenarios. Recently, transformer-based models have been introduced to the person re-identification field, achieving promising results. However, in person re-identification, most transformer-based models currently only use the features from the last stage for recognition. These high-level features lack fine-grained information and are not effective in addressing the feature misalignment issue. To overcome these challenges, we propose a novel MultiScale Features Fusion Swin Transformer (MSFF-Swin), which effectively integrates intermediate-layer and high-level features from the swin transformer. The inclusion of intermediate-layer features helps align the pedestrian features, addressing the feature misalignment caused by occlusion. Additionally, MSFF-Swin employs a multi-branch structure to downsample the intermediate-layer features, preserving the fine-grained information and effectively addressing the information loss brought by occlusion, thus improving the model's robustness in occluded environments. Experimental results on both occlusion and general re-identification datasets demonstrate that the proposed MSFF-Swin network exhibits excellent performance and robustness to occlusion.