Deep adversarial data augmentation with attribute guided for person re-identification

Person re-identification (Re-ID) is aimed at matching the identity class of pedestrian image across multiple different camera views. Most existing Re-ID methods rely on learning model from labeled pairwise training data. This leads to poor scalability and usability due to the lack of mass identity labeling of images for every camera pairs. In this paper, we address this problem by proposing a deep adversarial learning approach capable of generating images for person Re-ID. Specifically, we propose a deep adversarial data augmentation method with attribute (DADAA) which generates various person images by generative adversarial augmentation. The mid-level attribute information is integrated into the proposed DADAA, which is formulated as learning a one-to-many mapping from labeled source dataset to a large-scale target dataset for increasing data diversity against overfitting. Extensive comparative evaluations show that the DADAA method significantly improves the performance of person Re-ID and validate the superiority of this DADAA method over some state-of-the-art methods on Market-1501 and DukeMTMC-ReID.