FUsing Global and Semantic-Part Features with Multiple Granularities for Person Re-Identification

A multiple granularities method for person re-identification (re-ID) is proposed in this paper, which fuses global and semantic-part representations. A prior guided human parsing method is employed to parse a human body into precise basic semantic parts from low-resolution images, and multiple granularities are generated by recombining the adjacent basic semantic parts. Then, convolutional neural networks that seam-lessly unify the Softmax and TriHard losses are proposed to learn and fuse the global-level and the part-level features in different granularities. The proposed method not only extracts precise part-level features, but also incorporates gradual cues between part-level and global-level features to boost a high performance of person re-ID. Extensive experimental results show our proposed SPMG model achieves state-of-the-art performance on three common datasets .