Unsupervised learning of visual invariant features for person re-identification

Abstract

The recent successful methods of person re-identification (person Re-ID) involving deep learning have mostly adopted supervised learning algorithms, which require large amounts of manually labelled data to achieve good performance. However, there are two important unresolved problems, dataset annotation is an expensive and time-consuming process, and the performance of recognition model is seriously affected by visual change. In this paper, we primarily study an unsupervised method for learning visual invariant features using networks with temporal coherence for person Re-ID; this method exploits unlabelled data to learn expressions from video. In addition, we propose an unsupervised learning integration framework for pedestrian detection and person Re-ID for practical applications in natural scenarios. In order to prove the performance of the unsupervised person re-identification algorithm based on visual invariance features, the experimental results were verified on the iLIDS-VID, PRID2011 and MARS datasets, and a better performance of 57.5% (R-1) and 73.9% (R-5) was achieved on the iLIDS-VID and MARS datasets, respectively. The efficiency of the algorithm was validated by using BING + R-CNN as the pedestrian detector, and the person Re-ID system achieved a computation speed of 0.09s per frame on the PRW dataset.