Learning to disentangle scenes for person re-identification

There are many challenging problems in the person re-identification (ReID) task, such as occlusion and scale variation. Existing works usually tried to solve them by employing a one-branch network. This network needs be robust various problems, which makes this overburdened. paper proposes divide-and-conquer ReID task. For purpose, we employ several self-supervision operations simulate different handle each problem using networks. Concretely, use random erasing operation propose novel scaling generate new images with controllable characteristics. A general multi-branch network, including one master branch two servant branches, is introduced scenes. These branches learn collaboratively achieve perceptive abilities. In way, complex scenes task effectively disentangled, burden of relieved. The results from extensive experiments demonstrate that proposed method achieves state-of-the-art performances on three benchmarks occluded benchmarks. Ablation study also shows scheme significantly improve performance