Active learning with re-sampling for support vector machine in person re-identification

Abstract

Person re-identification is defined as to find the same person who re-occurred in a multi-camera surveillance system. A classifier for person re-identification may suffer from the imbalance dataset problem since the number of the targeted images is much less than irrelevant images. In this paper, we proposed over-sampling and under-sampling method for the active learning method for person re-identification. The sampling method is activated when the imbalance level of the training set is higher than a preset value during iteration of the active learning. The effect of the imbalance problem is reduced. Experimental results show the active learning method with the proposed re-sampling method scarifies the true negative rate to achieve higher true positive rate, which is more important in person re-identification.