A feature disentangling approach for person re-identification via self-supervised data augmentation

Abstract To address the problem of insufficient training data in person ReID, this paper proposes a data augmentation method based on image channels shuffling, by which a large volume of diversified training samples sharing similar edges can be produced. In the meantime, a soft label assignment strategy is designed to characterize the correlations between the original image and the generated counterparts. Furthermore, we design an encoder–decoder based learning structure for the person ReID task, where the encoder module tackles feature disentangling according to the introduced correlations, and the decoder module handles reconstruction using the combinations of decoupled features. Extensive experiments on four benchmark datasets demonstrate the effectiveness and robustness of the proposed method by attaining significant improvement over some state-of-the-art approaches. Source code is provided for reproducibility and it will be released at: https://github.com/flychen321/ID_aug_reid .