Person Re-identification via Structured Prediction

The goal of person re-identification (re-id) is to maintain the identity of an individual in diverse locations through different non-overlapping camera views. Re-id is fundamentally challenging because of appearance changes resulting from differing pose, illumination and camera calibration of the two views. Existing literature deals with the two-camera problem and proposes methods that seek to match a single image viewed in one camera to a gallery of images in the other. We propose structured prediction as a way to learn simultaneous matches across the two camera views. We deal with appearance changes in our prediction model through basis functions that encode co-occurrences of visual patterns in the two images. We develop locality sensitive co-occurrence measures as a way to incorporate semantically meaningful appearance changes. Empirical performance of our method on two benchmark re-id datasets, VIPeR [12] and CUHK Campus [38], achieves accuracy rates of 38.92% and 56.69%, at rank-1 on the so-called Cumulative Match Characteristic curves and beats the state-of-the-art results by 8.76% and 28.24%.