Zhuang Zijie ; Wei Longhui ; Xie Lingxi ; Zhang Tianyu ; Zhang Hengheng ; Wu Haozhe ; Ai Haizhou ; Tian Qi

Abstract

The fundamental difficulty in person re-identification (ReID) lies inlearning the correspondence among individual cameras. It strongly demandscostly inter-camera annotations, yet the trained models are not guaranteed totransfer well to previously unseen cameras. These problems significantly limitthe application of ReID. This paper rethinks the working mechanism ofconventional ReID approaches and puts forward a new solution. With an effectiveoperator named Camera-based Batch Normalization (CBN), we force the image dataof all cameras to fall onto the same subspace, so that the distribution gapbetween any camera pair is largely shrunk. This alignment brings two benefits.First, the trained model enjoys better abilities to generalize across scenarioswith unseen cameras as well as transfer across multiple training sets. Second,we can rely on intra-camera annotations, which have been undervalued before dueto the lack of cross-camera information, to achieve competitive ReIDperformance. Experiments on a wide range of ReID tasks demonstrate theeffectiveness of our approach. The code is available athttps://github.com/automan000/Camera-based-Person-ReID.