Multi-Scale Convolutional Network for Person Re-identification

In the last several years, methods with learning procedure held the state-of-the-art results for person re-identification (re-id) problem, especially the metric learning algorithm. Recently, with the success of deep learning methods on many computer vision tasks, researchers started to put their focuses on learning high-performance features. In this paper, we propose a method by fusing features learned from a multi-scale convolutional neural network and the traditional hand-crafted features, which improves the performance significantly. The Shinpuhkan2014dataset has been chosen as the training set, and we evaluate the performances of the proposed method on VIPeR, PRID and i-LIDS. Experiments show that our method outperforms the existing methods and even approaches the performances of the methods which have a training step on the testing sets.