Deep multi-instance learning for end-to-end person re-identification

Abstract

In this paper, we introduce a deep multi-instance learning framework to boost the instance-level person re-identification performance. Motivated by the observation of considerably dramatic and complex varieties of visual appearances in many current person re-identification datasets, we explicitly represent a deep feature representation learning method for the final person re-identification task. However, most public datasets for person re-identification are usually small, that usually make deep learning model suffer from seriously over-fitting problem. To alleviate this matter, we formulate the problem of person re-identification as a deep multi-instance learning (DMIL) task. More specifically, We build a novel end-to-end person re-identification system by unifying DMIL with the convolutional feature learning. For well capturing these intra-class diversities and inter-class ambiguities of input visual samples across cameras, a multi-scale convolutional feature learning method is proposed by optimizing the Contrastive Loss function. Comprehensive evaluations over three public benchmark datasets (including VIPeR, ETHZ, and CUHK01 datasets) well demonstrate the encouraging performance of our proposed person re-identification framework on small datasets.