Multi-task Learning for Person Re-identification

Person re-identification is a hot topic due to its huge application potentials. Siamese network is a good method to learn feature representation in verification tasks and has been used in previous person re-identification research, but hard to convergence during training process. This paper presents a multi-task learning pipeline including Siamese loss for learning deep feature representations of people appearance. Firstly, we point out the defects of training a convolutional neural network (CNN) only with Siamese loss which is usually used for person re-identification. Secondly, a multi-task CNN for person re-identification combing the Softmax loss with Siameses loss is proposed. Finally, some experiments are carried out to test the performance of proposed multi-task person appearance learning pipeline. Experiments on various pedestrian dataset shows the effectiveness of our pipeline. Our method outperforms state-of-the-art person re-identification methods in some public datasets.