Compact Triplet Loss for person re-identification in camera sensor networks

The triplet loss in deep learning has achieved promising results for person re-identification (re-ID) in camera sensor networks. However, it neglects the relationship among pedestrian images captured from different sensors, which results in a relatively large intra-class variation. In this paper, we propose a novel loss function named Compact Triplet Loss (CTL) for training Convolutional Neural Networks (CNNs), which not only decreases the intra-class variation but also increases the inter-class variation to improve the generalization ability of person re-ID model. Specifically, CTL simultaneously considers three aspects for pedestrian representations. It pushes the pedestrian images to be closer to their corresponding centers and meanwhile forces different centers are away from each other. In addition, CTL forces the distance between the positive sample pair is smaller than that of the negative sample pair. Finally, we integrate the proposed CTL and the cross-entropy loss to perform multi-task learning. We evaluate the proposed method on Market1501, DukeMTMCreID and CUHK03, and the experimental results reveal our method exceeds other state-of-the-art methods by a large margin. (C) 2019 Elsevier B.V. All rights reserved.