Non-full multi-layer feature representations for person re-identification

Person re-identification(Re-ID) has attracted increasing attention in the field of computer vision due to its great significance for the potential real-world applications. Profited from the success of convolutional neural networks(CNNs), existing multi-layer approaches leverage different scales of convolutional layers to learn more discriminative features, improving the Re-ID performance to some extent. However, these methods do not further explore whether all the scales of convolutional layers are positive for person re-identification. In this work, we propose a novel non-full multi-layer(NFML) network, which can jointly learn discriminative feature embeddings from positive multiple layers with the manner of combining global and local cues. Moreover, considering few works focus on how to effectively handle the feature maps, a simple yet effective feature progressing module named Pooling Batch Normalization(PBN), consisting of pooling, reduction and batch normalization operations, is introduced to optimize the model structure and further improve the Re-ID performance. Results on three mainstream benchmark datasets Market-1501, DukeMTMC-reID and CUHK03 demonstrate that our method can significantly boost the performances, outperforming the state-of-the-art methods.