On the Optimization of Multi-branch Cooperative OSNet for Person Re-identification

OSNet is an effective lightweight neural network architecture, which has attracted attention in the field of person re-identification duo to its excellent performance. Recently, we have proposed a multi-branch cooperative OSNet network based on OSNet, termed BC-OSNet, which performs significantly better than OSNet. In this paper, we study the further optimization of BC-OSNet with some adjustments on various micro-structures, including generalized-mean pooling, continuous Gaussian Dropout, attention modules of Batch DropBlock (BDB)/Relation-Aware Global Attention(RGA), etc. Experimental results show that the optimized BC-OSNet achieves 89.9%, 82.1%, 84.2%, and 81.5% mAP on the four pedestrian re-identification datasets, including Market1501, Duke, CUHK03_Labeled, and CUHK03_Detected, respectively. This means that the optimized BC-OSNet surpasses BC-OSNet about 0.6%, 1.4%, 1.1% and 1.7% in mAP for these datasets.