Abstract

Person re-identification across cameras remains a verychallenging problem, especially when there are no over-lapping fields of view between cameras. In this paper,we present a novel multi-channel parts-based convolution-al neural network (CNN) model under the triplet frameworkfor person re-identification. Specifically, the proposed CNNmodel consists of multiple channels to jointly learn both theglobal full-body and local body-parts features of the inputpersons. The CNN model is trained by an improved tripletloss function that serves to pull the instances of the sameperson closer, and at the same time push the instances be-longing to different persons farther from each other in thelearned feature space. Extensive comparative evaluation-s demonstrate that our proposed method significantly out-performs many state-of-the-art approaches, including bothtraditional and deep network-based ones, on the challeng-ing i-LIDS, VIPeR, PRID2011 and CUHK01 datasets.