Efficient Structure Search for Person Re-identification

Person re-identification (re-ID) has been a research hotspot in the field of computer vision, and in recent years, with the development of deep learning, combined re-ID with deep learning has achieved promising performance. Pedestrian features, however, can be easily affected by various factors such as illumination, occlusion, and viewpoint changes, which cannot be fully expressed. In this paper, we use ResNet-50 network as the backbone network for feature extraction. Firstly, we embed a lightweight attention mechanism to the residual blocks in it to enhance the feature information extraction. Secondly, the training techniques of random erasure using Cutout, downsampling using generalized mean pooling, and adding ArcFace loss improve the performance of the model. Finally, the features are processed by convolution, batch normalization, and Relu activation to enhance the feature representation of the model. Through extensive experiments on Market1501 and DukeMTMC-ReID datasets, Rank1 is improved by 5.9% and 5.5%, and mAP is improved by 9.8% and 6.8%, respectively, and the results show that the improved model can achieve better re-ID performance.