Pair-dependent representations for Same-View Vehicle recognition

Deep metric learning approaches for object instance recognition struggle with generalization. To infer if two images are of the same instance, a representation is computed for each image, independently of the other. This is insufficient for reliably finding distinctive fine-grained details. In this paper, we focus on the problem of same-view vehicle recognition. We propose a keypoint matching-based method that combines pair-independent and pair-dependent information to predict whether two images are instances of the same object. We propose an evaluation protocol focused on generalization to unseen types of vehicles. Extensive experiments show that the proposed method generalizes better to unseen vehicle types than the state of the art.