Equidistant distribution loss for person re-identification

Abstract

Person re-identification (re-id) is a technology that matches images of a certain pedestrian among non-overlapping cameras. Recently, many approaches based on deep learning have been studied comprehensively. Most of these methods try to learn discriminative pedestrian features via elaborated network architectures and one or multi loss functions to maximize inter-class separability and intra-class compactness. However, the class imbalance characteristics from the datasets are ignored, and the discriminative information cannot be completely exploited. In this paper, we propose a novel loss function termed as Equidistant Distribution Loss (EDL) to improve the imbalance problem that exists in re-id tasks. Specifically, we first normalize the learned features in the embedding layer and the weights in the last fully connected layer, so that both the features and the weights can be projected on a hypersphere space. After that, we directly impose an equidistance constraint among the weights to guide the learned features spread uniformly in the sphere space. The proposed EDL can effectively address the local squeeze or imbalance problem caused by imbalanced samples, making full use of feature space and profiting the performance. Extensive experiments on re-id datasets, including Market-1501, DukeMTMC-reID are conducted to demonstrate the preferable performance of the proposed method.