Discrimination Power of Body Parts in Person Re-identification: An Evaluation by Histogram Based Warping Function

Abstract

Person re-identification is a well-known technique for recognizing a person, who is continuously observed by nonoverlapping cameras in a wide area surveillance. Body part-based re-identification gives a higher level of importance to certain parts of body instead of giving equal importance to whole body. This is mainly due to the fact that certain body parts can be distinguishable better than the others in presence of variable viewing positions, occlusions, illuminations and resolutions of the scenes. This paper focuses on the evaluation of the discrimination power of three body parts, viz., head, torso, and leg for their potential use in fusion-based re-identification technique. In particular, the transformation of features between two body parts is conducted in a warp function space that consists of positive cost for same pair of targets and negative cost for different pair of targets. The support vector machine-based classifier is used to discriminate these two types of cost and the results are evaluated in terms of correct recognition for a given false alarm rate. Experiments are conducted on two publicly available databases, namely, CAVIAR4ReID and CUHK01. The results conclude that torso is the most important body part in the problem of person re-identification.