Person Re-Identification by Scalable Manifold Ranking with User Side Information Enhancement in Surveillance System

摘要

The goal of person reidentification (Re-ID) is to identify a given pedestrian from a network of nonoverlapping surveillance cameras. Most existing works conventionally rely on labelled pairwise data to learn a task specific distance metric for ranking. The value of unlabelled gallery instances is generally overlooked. On the other hands, Manifold Ranking (MR) has been successfully applied to Person ReID due to its ability to discover underlying geometrical structure of dataset given the query data. However, one problem of MR is that it cannot be scalable to large-scale data. To solve this problem, we develop a new scalable manifold ranking method for Person ReID by adopting an anchor graph for construction, where each data point is first to find the k neighbors of anchor points, and then the graph is constructed by the inner product of coefficients between the data points and anchor points. As a result, the proposed method can well handle a database with large-scale person images and do the online retrieval in a short time. Extensive evaluation is conducted on three benchmark datasets.