Domain Feature Alignment Based on Complement Foreground for Cross-domain Pedestrian Re-identification

Although significant progress has been made in supervised pedestrian re-identification (re-id), it is still challenging to extend the re-id model to new scenes due to the huge domain gap. Getting rid of domain-specific features and aligning domain features are the ideas to solve cross-domain problems. However, for the former, the existing schemes are often not thorough enough to get rid of the unique characteristics of the domain, and for the latter, domain feature alignment schemes try to align the background and other information that should not be aligned, so they fail to achieve ideal results. In this paper, we propose domain alignment based on complementary foreground. The scheme strips off background clutter thoroughly with the help of mask, and uses pedestrian attribute information and control factor to reduce the influence of background loss caused by masks. Finally, we measure the distribution difference between the two domains with maximum mean difference, and make the feature distribution farther or closer according to the different feature levels. Experiments on mainstream data show that our scheme is 13.9% higher than the benchmark, and extensive ablation experiments prove the effectiveness of our design.