Multi-Branch Network with Dynamically Matching Algorithm in Person Re-Identification

In person re-identification problem, learning distinctive features to differentiate one person from the others is one of the key factors to improve the result. In this paper, we propose the Multi-Branch Network with Dynamically Matching (MNDM) algorithm, which has a multi-branch deep network architecture consisting of three branches: one for learning global features, and two for local features. In the one branch learning local features, where the bounding box is split into horizontal stripes, the model applies a dynamically matching algorithm to efficiently matching these parts addressing the misalignment issue caused by imperfect detection bounding boxes. Several experiments are conducted on Market1501, CUHK03 and DukeMTMC. All the results that demonstrate the proposed model significantly outperforms the previous state-of-the-art.