Improving person re-identification by multi-task learning

We propose a novel Multi-Task Learning Network (MTNET) with four different subtasks for person re-identification mission. At the same time, the attribute recognition mission can be implemented by the same network. We achieve multi-mission by integrating four subtasks, such as identity identification, identity verification, attribute identification, attribute verification. Identity loss and attribute loss can provide complementary information on a different perspective by integrating multi-context information. Identity focuses on the overall contour and appearance, while attribute focuses on local aspects and dresses of one person. Identification loss and verification loss are used to optimize the distance of samples. Identification loss used to construct a robust category space, while verification loss used to optimize the space by minimizing the distance between similar images, and maximizing the distance between dissimilar images. Moreover, an effective verification loss named constraint contrast verification (CCV) is proposed to restrict the distance between feature pair to a foreseeable range that ensures the network has better convergence. The MTNet is an end-to-end deep learning framework, all the parameters and losses can be jointly optimized. We evaluate our approach with the state-of-the-art methods on two famous dataset Market1501 and DukeMTMC-reID. Experiments demonstrate that our MTNet achieves the very competitive results.