Learning fused features with parallel training for person re-identification

摘要

The single-image representation (SIR) matching and cross-image representation (CIR) classification are two significant solutions in the person re-identification (Re-ID) task. Combining the two categories of methods has been regarded as an effective solution to improve discriminative performances in Re-ID. Previous combination methods mainly focus on fusing the SIR learning and CIR learning losses, representing the cross-image features with limitations and not available to different network structures for feature learning. This paper proposed an efficient joint SIR and CIR learning strategy — cross-image features fusion strategy (CFFS), which fuses cross-image information through parallel training. Precisely, CFFS consists of one shared sub-network and two branches, one branch for learning single-image feature and the other for fusing cross-image features. CFFS requires pairwise data parallelism training for each identity to learn the CIR, applied to metric learning. Therefore, the cross-image features would be fused better, and the performance of Re-ID would be improved. Experiments on Market-1501, DukeMTMC-reid, and CUHK03 datasets show that the proposed method can achieve favorable accuracy while compared with state-of-the-arts.