Person re-identification via adaptive verification loss

Abstract

Person re-identification (re-ID) is the task of associating the image patches of the identical person across disjoint cameras, which is a very challenging topic in computer vision. Recently, with the advance of deep learning techniques and the public of huge datasets, person re-ID has achieved a great success. However, an obvious but ignored issue is that the sample pairs are biased, where the negative sample pairs are far more than the positive ones. In this paper, we develop an adaptive verification loss, termed as ADV-Loss to handle the imbalance of sample pairs. Our ADV-Loss empowers the popular verification loss with the flexibility of adaptively addressed the hard triple units. Specifically, we learn a re-weighted strategy from the triplet loss to balance the sample pairs. According to the objective values of triple loss, the hard triple units will be endowed with larger weights, while the less important triple units are de-emphasized or simply dropped. Thus, ADV-Loss, on the basis of the cross-entropy loss, handles the biased issue by only holding the equal number but informative sample pairs. Moreover, theoretical analysis of our designed ADV-Loss is provided to guarantee the important sample pairs not to be despised during the backpropagation. Ablation analysis verifies the potential efficacy of the proposed loss. Moreover, experiments of person re-ID on three large-scale benchmark datasets show that two popular deep networks with the proposed loss achieve satisfactory performance against several well-established person re-ID methods.