Multi-Scale Deformable Convolutional Network for Visible-Infrared Vehicle Re-Identification

Vehicle re-identification (Vehicle-ReID) has become a focal point in intelligent transportation systems in recent years. Conventional vehicle re-identification studies are primarily relying on visible single-modal data, neglecting cross-modal performance, which limits their application in complex traffic scenarios. This paper proposes a Multi-Scale Deformable Convolutional Network (MSDCN), which learns feature representations of images through linear deformable convolutional kernels of different scales. To assess the model's performance more thoroughly, we collected an open-source visible-infrared target tracking dataset and organized it into a visible-infrared cross-modal vehicle re-identification dataset, Vehicle-138. This dataset contains images of vehicles from 138 identities, including 43,304 visible images and 43,328 infrared images. The Multi-Scale Deformable Convolutional Network outperforms other advanced methods on the Vehicle-138 dataset, demonstrating stronger cross-modal recognition capabilities.