Unsupervised Cross-domain Person re-Identification by Deep Clustering and Instance Learning

Abstract

Cross-domain unsupervised person re-identification (Re-id) has become more and more popular due to cost of labeled images. However, because of the large differences between two different domains in lighting, background, and so on, cross-domain unsupervised person Re-id is still a very challenging task. Most of the current mainstream methods utilize the labeled source domain data and unlabeled target domain data to train a CNN network, and apply it to the un- labeled target domain. In this paper, we consider the intra-domain variations of the target domain and propose a deep clustering and instance learning (DCIL) approach. Our method considers two factors simultaneously: (1) instance invariance. We use sample memory module to save features of each class, and enforce each person to be close to its corresponding instance; (2) instance similarity. We use clustering to obtain the pseudo-labels for the unlabeled domain instances, and make each person be close to its similar instance so as to minimize the wrong pseudo-labels. We propose a clustering repelled loss to learn discriminative features for the unlabeled data while considering the above two factors. Extensive experiments on benchmark datasets demonstrate the superiority of our method for unsupervised cross-domain person Re-id.