A Pedestrian Re-Identification Network Based on Multi-Granularity Feature Extraction

Pedestrian re-identification aims to distinguish and recognize a particular pedestrian within a set of images taken by various cameras. Previous research has demonstrated that simple, global and local features may not yield optimal results for the model. In this study, an end-to-end pedestrian recognition network is enhanced by incorporating multi-granularity pedestrian features. This network comprises multiple deep network modules, specifically, two global feature branch modules and two local feature branch modules. To augment the model's feature extraction capability, we have integrated an improved parameterless attention module into the baseline model, resulting in significant performance improvements. We have conducted comprehensive experiments on well-established evaluation datasets, including Market-1501, DukeMTMC ReID, and MSMT17, and the results have empirically demonstrated advanced performance that surpasses most current methods. Taking the Market-1501 dataset as an illustration, our approach demonstrated a mean Average Precision (mAP) of 94.0% and a Rank-1 accuracy of 95.8% in single query mode, with the application of reordering techniques.