Person re-identification using compound descriptor and invisible region replacement

Abstract

In this paper we proposed person re-identification algorithm using compound descriptor that includes global and local features for the top, middle, and bottom of the person figure. Local areas are formed based on the person figure key points coordinates. If there are not enough visible points, the area is recognized as invisible and feature vector corresponding component is replaced by an average value for the k-nearest neighbors. Testing was performed on datasets for re-identification Market-1501, DukeMTMC-ReID, MSMT17, PolReID1077. Our algorithm allows us to increase accuracy re-identification for metric Rank1 by 8 - 51% and for metric mAP by 28 - 97% relative to the baseline