Two Stream Pose Guided Network for Vehicle Re-identification

Vehicle Re-Identification is the task of finding images of the same vehicle with different views across a surveillance camera network, which is a very beneficial yet challenging task. Huge intra-class differences and small inter-class difference makes this task hard to tackle. Appearance-based information is utilized in this paper to cope with vehicle re-identification problem; we have proposed a deep learning technique by incorporating poses of vehicles generated by pose estimation network and visual information. When query image is given, the two-stream network generates a feature embedding by concatenating pose feature from pose network. Extensive experiments are done on two of the benchmark datasets of vehicle re-identification VeRi-776 and VehicleID. Experimental results are supporting the competitiveness of the proposed method with recent state-of-the-art methods.