Identity Preserving Generative Adversarial Network for Cross-Domain Person Re-Identification

In this paper, we study the domain adaptive person re-identification(re-ID) problem: train a re-ID model on the labeled source domain and test it on the unlabeled target domain. It's known challenging due to the feature distribution bias between the source domain and target domain. The previous methods directly reduce the bias by image-to-image style translation between the source and the target domain in an unsupervised manner. However, these methods only consider the rough bias between the source domain and the target domain but neglect the detailed bias between the source domain and the target camera domains (divided by camera views), which contain critical factors influencing the testing performance of re-ID model. In this work, we particularly focus on the bias between the source domain and the target camera domains. To overcome this problem, a multi-domain image-to-image translation network, termed Identity Preserving Generative Adversarial Network (IPGAN) is proposed to learn the mapping relationship between the source domain and the target camera domains. IPGAN can translate the styles of images from the source domain to the target camera domains and generate many images with styles of target camera domains. Then the re-ID model is trained with the translated images generated by IPGAN. During the training of the re-ID model, we aim to learn the discriminative feature. We design and train a novel re-ID model, termed IBN-reID, in which Instance and Batch Normalization block (IBN-block) are introduced. Experimental results on Market-1501, DukeMTMC-reID and MSMT17 show that the images generated by IPGAN are more suitable for cross-domain re-ID. Very competitive re-ID accuracy is achieved by our method.