Common visual part alignment for vehicle re-identification

Vehicle re-identification (Re-ID) aims at identification of images with the same vehicle ID in a cross-camera scenario. Prior methods mainly tend to fuse global features with local features. However, huge divergence in viewpoints among cameras leads to inconsistency of local features between query and gallery images, which causes performance drop. Recently, alignment-based methods have been proposed to handle this problem. However, previous alignment-based methods rely on the detection accuracy of vehicle component excessively. To solve this problem, a common visual part alignment (CVPA) model to align local features between image pairs with a few visual key points in image pairs is proposed. The main contributions of CVPA include: 1) CVPA utilizes common visual feature map generated by key points to align local features without fine-grain annotations. 2) Using key-point coordinates based on Gaussian smoothing can reduce the negative influence of key prediction deviation, which leads CVPA to have more flexibility and better adaptability. 3) By proposing orientation-quadruplet loss to mine hard samples, CVPA is more robust against vehicle orientation variation. Extensive experiments show the superior performance of CVPA.