Two-stream person re-identification with multi-task deep neural networks

Person re-identification (re-id) with images is very useful in video surveillance to find specific targets. However, it is challenging due to the complex variations of human poses, camera viewpoints, lighting, occlusion, resolution, background clutter and so on. The key to tackle this problem is how to represent the body and match these representations among frames. Current methods usually use the features of the whole bodies, and the performance may be reduced because of part invisibility. To solve this problem, we propose a two-stream strategy to use parts and bodies simultaneously. It utilizes a multi-task learning framework with deep neural networks (DNNs). Part detection and body recognition are performed as two tasks, and the features are extracted by two DNNs. The features are connected to multi-task learning to compute the mapping model from features to identifications. With this model, re-id can be achieved. Experimental results on a challenging task show the effectiveness of the proposed method.