Multi-scale Local Implicit Keypoint Descriptor for Keypoint Matching

We investigate the potential of multi-scale descriptors which has been under-explored in the existing literature. At the pixel level, we propose utilizing both coarse and fine-grained descriptors and present a scale-aware method of negative sampling, which trains descriptors at different scales in a complementary manner, thereby improving their discriminative power. For sub-pixel level descriptors, we also propose adopting coordinate-based implicit modeling and learning the non-linearity of local descriptors on continuous-domain coordinates. Our experiments show that the proposed method achieves state-of-the-art performance on various tasks, i.e., image matching, relative pose estimation, and visual localization.