Cross-Camera Prototype Learning for Intra-camera Supervised Person Re-identification

Person Re-Identification (ReID) aims at retrieving images of the specific pedestrian across disjoint cameras. However, the annotations are extremely costly as the number of cameras increases, which derives a new setting named Intra-Camera Supervision (ICS) ReID. ICS assumes that identity labels are independently annotated within each camera, while no cross-camera identity association is available. Previous ICS methods focus on connecting the inter-camera instances that are likely to be the same pedestrian, whereas fails to exploit the so far untapped yet informative supervision, i.e., 'the cross-camera prototype relations'. In this paper, we propose the novel Cross-Camera Prototype Learning (CCPL) method to tackle this issue. Firstly, we ensure identities to be discriminative and associated with corresponding intra-camera prototypes, which can be considered as the semantic representations for each local identity. Besides, we claim that the distance between the same inter-camera prototypes is inevitably large, due to the variances of different cameras in views, lights, backgrounds etc. To that end, we propose the Camera-invariant Prototype Alignment (CPA) module, which preserves the cross-camera prototype relations by explicitly pulling together the same inter-camera prototypes and pushing away the different ones. Last but not least, we also introduce the inter-camera prototype pulling loss to constrain the same prototypes as close as possible. Extensive experiments on three benchmarks show the superiority of our method.