Efficient but lightweight network for vehicle re-identification with center-constraint loss

Vehicle re-identification aims to retrieve the target vehicle from the image gallery quickly and accurately. Vehicle re-identification with deep learning has achieved considerable performance. However, most popular methods need construct complex network, which increases the calculation of network and difficulty of training. To balance performance and complexity, an efficient but lightweight network is proposed in this work. The designed network utilizes the global branch and the mask branch to extract the feature. The former can extract global feature efficiently. The latter can remove the changeable background based on the proposed mask-mapping module, which can map mask to feature map and adjust feature map dynamically. And then, the features from the original image and the mask-mapping module are fused to generate the final feature for vehicle re-identification. Besides, a novel center-constraint triplet loss is designed to optimize the proposed network and excavate more discriminate feature. Different from triplet loss, the proposed loss can consider more extra samples and constrain the center from positive sample set as well as negative sample set. To enhance the difference between hard samples and simple samples, an unequal weight strategy is embedded in this loss. The proposed method achieves 78.7% mAP with 95.4% Rank-1 on VeRi-776, and 84.3%, 80.7%, and 80.1% Rank-1 on three subsets from VehicleID, which demonstrates the effectiveness of the proposed method.