Person Re-identification Based on Improved Transformer and Multi-scale, Multi- granularity Feature Learning

This is the authors' abstract. We don't add key points for this paper.

Person re-identification refers to the use of computer vision techniques to identify specific individuals in images captured by surveillance systems. However, this process is challenged by factors like variations in viewpoint, pose, occlusions, and other environmental conditions. Current pedestrian re-identification methods based on local features often suffer from limited discriminative power in feature extraction, which ultimately results in lower re-identification accuracy. For efficient use of feature information, we propose an improved Transformer and multi-scale, multi-granularity fusion network. Firstly, local multi-granularity branch is introduced to extract semantic information at various granularities, which enhances the richness of features by exploring less significant local characteristics of pedestrians. In addition, the improved Transformer can better focus on key regions of the image and incorporates larger-scale image information. Finally, we use triplet and softmax to calculate loss for different branches. We conducted extensive experiments on three datasets (CUHK03, Market1501, DukeMTMC-reID), and the results demonstrate that our model outperforms many recent state-of-the-art methods in terms of accuracy.