Abstract: In order to address the problem of insufficient discriminative features for pedestrian re-identification extracted by exis-ting convolutional neural network,a novel multi-scale multi-granularity feature learning for pedestrian re-identification method is proposed.In the training phase,the method extracts multi-scale features at different stages of the convolutional neural network,and then blocks and pools these feature maps to obtain multi-granularity features containing global and local features,uses uncertainty to weight Softmax loss and triples loss and to supervise training process on feature vectors.In the inference phase,the obtained multi-scale multi-granularity features are concatenated,and finally the concatenated features are used to perform similarity matching in the gallery.Experiments on the Market-1501 and DukeMTMC-ReID datasets show that the proposed method improves the Rank-1 evaluation index by 4.3% and 3.6%,respectively,compared with the benchmark network ResNet-50,and improves the mAP evaluation index respectively 6.2% and 6.6%.The results show that the proposed method can enhance the discrimination of extracted features and improve the performance of pedestrian re-identification.

Key words: Convolutional neural network, Machine vision, Multi-granularity features, Multi-scale features, Pedestrian re-identification