Unsupervised Person Re-identification via Differentiated Color Perception Learning

Unsupervised person re-identification (re-ID) encounters two key problems. One is the lack of label annotation in the target domain, and the other is the domain gap between different cameras. They are addressed in this paper based on the framework of pseudo label estimation-based re-ID. For the former issue, we firstly take advantage of HSV color space to design a novel data augmentation strategy, with which generated samples with controllable color components can be obtained. We then construct self-contained feature-level supervision on the augmented samples. For the latter issue, we design an explicit camera-related correction term to relieve the negative effects of camera differences, rather than suppressing the sensibility of the model to views through adversarial learning. Therefore, our model can better hold the perception to pedestrian appearance. Comprehensive experiments on three benchmark datasets have verified the superiority of our approach. Specifically, our method achieved over 1.0% performance improvement in terms of mAP compared to state-of-the-art methods. Code is available at https://github.com/flychen321/DCPL.