2003 Cvpr Homography

CVPR ‘20 (Seattle): For the classic homography estimation problem with outliers, popular methods based on RANSAC or deep networks often have costly computation and little theoretical guarantee. By viewing it as robustly fitting a 1/3/6-dimensional nullspace on homographic/epipolar embeddings, we propose a non-convex non-smooth \(\ell^1\) optimization, give theoretical analysis on the nullspace and global optimality, and show state-of-the-art performance!