Closed and Open-World Person Re-Identification and Verification

In recent years, remarkable breakthrough has been achieved in person re-identification (Re-ID). However most methods are only tested in the closed-world setting where the probe person is assumed to be one of the gallery people. In this paper, we tackle a more realistic problem, open-world Re-ID, which requires to find out whether the probe person is among the gallery or not, and if so, who he is. We formulate open-world Re-ID as a verification task which aim is to determine whether two sets of images represent the same person or two distinct people. We propose to learn a linear transformation of the features so that the distance between features of the same person are below a threshold and that of distinct people are above that same threshold. Tested on iLIDS-VID and PRID2011 datasets, for closed and open-world scenarios, our COPReV method shows promising results.