Online learning on incremental distance metric for person re-identification

Person re-identification is to match persons appearing across non-overlapping cameras. The matching is challenging due to visual ambiguities and disparities of human bodies. Most previous distance metrics are learned by off-line and supervised approaches. However, they are not practical in real-world applications in which online data comes in without any label. In this paper, a novel online learning approach on incremental distance metric, OL-IDM, is proposed. The approach firstly modifies Self-Organizing Incremental Neural Network (SOINN) using Mahalanobis distance metric to cluster incoming data into neural nodes. Such metric maximizes the likelihood of a true image pair matches with a smaller distance than that of a wrong matched pair. Second, an algorithm for construction of incremental training sets is put forward. Then a distance metric learning algorithm called Keep It Simple and Straightforward Metric (KISSME) trains on the incremental training sets in order to obtain a better distance metric for the neural network. Aforesaid procedures are validated on three large person re-identification datasets and experimental results show the proposed approach´s competitive performance to state-of-the-art supervised methods and self-adaption to real-world data.