Integration Convolutional Neural Network for Person Re-Identification in Camera Networks

In this paper, we propose a novel deep model named integration convolutional neural network (ICNN) for person re-identification in camera networks, which jointly learns global and local features in a unified framework. To this end, the proposed ICNN simultaneously applies two kinds of loss functions. Specifically, we propose the soft triplet loss to learn global features which automatically adjusts the margin threshold within one batch. The soft triplet loss could alleviate the difficult in tuning parameters and therefore learns discriminative global features. In order to avoid the part misalignment problem, we learn latent local features by conducting local horizontal average pooling on the convolutional maps. Afterward, we implement the identification task on each local feature. We concatenate global and local features using a weighted strategy to present the pedestrian images. We evaluate the proposed ICNN on three large-scale databases. Our method achieves rank-1 accuracy of 92.13% on Market 1501, 61.4% on CUHK03 and 85.3% on DukeMTMC-reID, and the results outperform the state-of-the-art methods.