Self-training with one-shot stepwise learning method for person re-identification

摘要

Person re-identification (Re-ID) aims at identifying the same person across multiple non-overlapping camera views. A number of existing methods have been presented for this task in a fully-supervised manner that requires a large amount of training annotations. However, obtaining high quality labels is extremely time consuming and expensive. In this article, we focus on the semi-supervised person Re-ID and propose a one-shot stepwise learning method to address the above issue. It exploits only one labeled data along with additional unlabeled samples to gradually but steadily improving the discriminative capability of the feature representation. Specifically, we first construct labeled data portion to train Re-ID model. Then we fine-tune the overall system by the following two steps iteratively: (1) assigning the estimated labels to the unlabeled portion; (2) updating the network parameters according to the selected data. During the propagation process, different from conventional sampling method, we propose a novel dynamic sampling strategy to enlarge the pseudo-labeled subset step by step to make the pseudo labels more reliable. On Market-1501, DukeMTMC-ReID and MARS datasets, we conducted extensively experiments to demonstrate that our proposed method contributes indispensably and achieves a very competitive Re-ID performance.