Person re-identification using Hybrid Task Convolutional Neural Network in camera sensor networks

Shuang Liu, Wenmin Huang, Zhong Zhang

Abstract: This paper proposes a new framework called Hybrid Task Convolutional Neural Network (HTCNN) which combines the advantages of ranking and classification tasks for person re-identification (re-ID) in camera sensor networks. As for the ranking task, we propose Weighted Triplet Loss (WTL) to optimize global features of pedestrians, and meanwhile WTL emphasizes the foreground of pedestrian image and weakens the background in order to enhance the feature discrimination. As for the classification task, we evenly divide the convolutional activation map into several horizontal parts and utilize average pooling to obtain local features of pedestrians. We evaluate our method on public person re-ID datasets, and the results indicate HTCNN exceeds the state-of-the-art re-ID methods.