I wonder why colmap is so slow at feature matching. Naively I can see how it would be O(n^2), but surely you can do better than that without needing silly hardware?