Refining pseudo labels for unsupervised Domain Adaptive Re-Identification

摘要

Our paper focuses on the topic of unsupervised domain adaptation for person re-identification (UDA re-ID) in an end-to-end framework. Currently, most existing methods tackle the problem by mining pseudo labels, and they rely heavily on the reliability of these pseudo labels. However, the generation of these labels mainly depends on clustering-based algorithms, which may inevitably introduce noise into the framework. Such noise substantially hinders the capability of the model to further improve feature representations in the target domain. We called these clustering-based labels HARD pseudo labels. To decrease the impacts of these noisy hard labels, in this paper, we directly consider the relative relationships of sample pair distances as another kind of pseudo label instead of clustering results, called SOFT pseudo labels. Furthermore, considering that the constructed relative relationships are not always correct and the wrong relationships may cause negative effects, it is not appropriate to assign the same weight to each sample pair. To alleviate the negative influence caused by tiny distances and strengthen the positive effect due to large distances, we assign different weights according to the different distances to adjust these relationships. By integrating the HARD labels and SOFT labels, the proposed method achieves considerable improvements on several popular person re-ID datasets, e.g, Market1501-to-Duke, Duke-to-Market1501, Market-to-MSMT17 and Duke-to-MSMT17 UDA tasks.