Camera Style and Identity Disentangling Network for Person Re-identification

Ruochen Zheng, Lerenhan Li, Chuchu Han, Changxin Gao, Nong Sang

Abstract: Camera style (camstyle) is a main factor that affects the performance of person re-identification (ReID). In the past years, existing works mainly exploit implicit solutions from the inputs by designing some strong constraints. However, these methods cannot consistently work as the camstyle still exists in the inputs as well as in the intermediate features. To address this problem, we propose a Camstyle-Identity Disentangling (CID) network for person ReID. More specifically, we disentangle the ID feature and camstyle feature in the latent space. In order to disentangle the features successfully, we present a Camstyle Shuffling and Retraining (CSR) scheme to generate more ID-preserved and camstyle variation samples for training. The proposed scheme ensures the success of disentangling and is able to eliminate the camstyle features in the backbone during the training process. Numerous experimental results on the Market-1501 and DukeMTMC-reID datasets demonstrate that our network can effectively disentangle the features and facilitate the person ReID networks.