Coarse-Fine Convolutional Neural Network for Person Re-Identification in Camera Sensor Networks

Abstract

In this paper, we present a novel deep model named coarse-fine convolutional neural network (CFCNN) for person re-identification in camera sensor networks, which jointly learns global and multi-scale local features simultaneously. To this end, we design the CFCNN as a multi-branch network, which is composed of one coarse and two fine branches. Specifically, the global feature is learned from the coarse branch, and the two fine branches are developed to extract two kinds of local features with different scales. Afterward, each branch is followed by a classification loss to make the identity prediction. Finally, we obtain completed pedestrian representations via concatenating the learned global and all local features. We conduct a number of experiments to evaluate the effectiveness of the CFCNN on three datasets. The CFCNN achieves high rank-1 and mAP accuracy with 94.0%/81.2%, 64.6%/58.4%, and 85.7%/72.4% on Market-1501, CUHK03, and DukeMTMC-reID, respectively. These results significantly outperform the prior state-of-the-art methods.