A Multi-loss Deep Network Based on Pyramid Pooling for Person Re-identification

Person re-identification (re-ID) can be regarded as a retrieval problem. The challenge of this task is mainly lying in that people's appearance often undergoes dramatic changes across camera views due to severe occlusions, illumination, complex background clutter, and large pose variation. Recently, many deep learning approaches are successfully employed in person re-ID, which have achieved state-of-the-art performance in a period. However, it easily neglects the local discriminative details on the images and is not enough to cover discriminative information. Therefore, we employ ResNet50 as a basic network and propose a multi-loss siamese network with pyramid pooling structure for person re-ID. Our network can get a better discriminative representation and achieve a better performance on two large public datasets including Market1501 and CUHK03. We also report competitive accuracy compared with previous approach, which demonstrates the effectiveness of proposed method.