Learning hierarchical representation using Siamese Convolution Neural Network for human re-identification

Abstract

Human re-identification is to match a pair of humans appearing in different cameras with non-overlapping views. However, in order to achieve this task, we need to overcome several challenges such as variations in lighting, viewpoint, pose and colour. In this paper, we propose a new approach for person re-identification in multi-camera networks by using a hierarchical structure with a Siamese Convolution Neural Network (SCNN). A set of human pairs is projected into the same feature subspace through a nonlinear transformation that is learned by using a convolution neural network. The learning process minimizes the loss function, which ensures that the similarity distance between positive pairs is less than lower threshold and the similarity distance between negative pairs is higher than upper threshold. Our experiment is achieved by using a small scale of dataset due to the computation time. Viewpoint Invariant Pedestrian Recognition (VIPeR) dataset is used in our experiment, since it is widely employed in this field. Initial results suggest that the proposed SCNN structure has good performance in people re-identification.