Transfusing Dense Matcher Experience into Sparse Keypoint Detectors in Weak-Texture Scenes

Sparse keypoint matching is essential for many vision tasks but often suffers in weak-texture scenes due to a lack of distinctive local features. We present a dense-to-sparse knowledge distillation framework that transfers multi-scale descriptors from a high-capacity dense matcher to a sparse keypoint detector. We align the teacher's feature pyramid via orthogonal projections and parameter-free channel attention, fuse them into a unified embedding, and use the teacher's match confidence to guide region selection. A composite loss enforces descriptor alignment, supervises keypoint confidence, and encourages homography-based consistency. On an indoor weak-texture benchmark, our distilled SiLK-D2S achieves up to 14.71% improvement in HEA and 4.17% gain in repeatability over the original SiLK, while maintaining real-time performance.