Self-supervised data augmentation for person re-identification

Although person re-identification (ReID) has been intensively studied over the past few years, the shortage of annotated training data stands as an obstacle for further performance improvement. To address this issue, many data augmentation methods have been successfully applied to person ReID, such as random scaling, flipping and cropping, which mainly operate on a single image, whilst overlooking the relationship between images. Recently, generative adversarial networks (GANs) have been widely used for data augmentation and smoothly migrated to person ReID. However, the cost of training GAN is expensive and the performance improvement is often limited. Moreover, all these methods focus on augmenting samples for existing IDs. This paper proposes a simple yet effective data augmentation strategy based on self-supervised learning to handle these problems, which consists of the offline ID augmentation that can generate new categories and the online instance augmentation. These two components are integrated i nto a unified framework for boosting the ReID performance. Furthermore, a novel Siamese-like network is developed for ReID in conjunction with the proposed data augmentation method. Extensive experiments on three benchmark datasets demonstrate that the proposed approach outperforms the state-of-the-art by a clear margin, which verifies the robustness and the effectiveness of our method. Code will be released at: https://github.com/flychen321/data_aug_reid. (C) 2020 Elsevier B.V. All rights reserved.