🤖 AI Summary
Existing superpixel neighborhood descriptors struggle to effectively model boundary contours, leading to insufficient robustness in cross-scale matching. This paper proposes a multi-scale superblock matching framework featuring a novel appearance-geometry coupled dual-superpixel descriptor, enabling semantically consistent cross-scale superblock representation. The method integrates superpixel segmentation, multi-scale pyramid sampling, dual-stream feature embedding, and differentiable matching optimization. Evaluated on benchmarks including HPatches, it significantly improves matching accuracy and recall under large viewpoint changes, strong illumination variations, and motion blur, while achieving higher inference efficiency than SIFT+RANSAC. Key contributions are: (1) a boundary-aware dual-stream superpixel descriptor mechanism that jointly encodes appearance and geometric structure; and (2) an end-to-end differentiable multi-scale matching paradigm that unifies feature learning and correspondence optimization.
