Channel Splitting Attention for Enhanced Person Re-Identification: A CSA-TOPDB Approach

This study introduces CSA-TOPDB, a network designed to enhance the performance of person re-identification. Building upon the TOPDB architecture with a ResNet-50 backbone, CSA-TOPDB integrates a channel attention mechanism that facilitates the extraction of more discriminative and detailed features. The key innovation lies in the channel-wise splitting of feature maps at different layers, followed by the application of channel attention to each segment. Initially, the Squeeze-and-Excitation (SE) module is used, later replaced by the more efficient Efficient Channel Attention (ECA) module. The inclusion of global processing and regularization branches further improves the network's robustness and generalization. Experimental results on four benchmark datasets demonstrate significant improvements over baseline techniques, with increases of 1.4% in mAP and 0.5% in R1 accuracy on the Market1501 dataset, and even greater enhancements on the DukeMTMC dataset. These results validate the effectiveness of the proposed channel splitting attention approach for person re-identification. Code available at https://github.com/hamed-ab4/CSA_TOPDB