Fixing the locally optimized RANSAC - full experimental evaluation

Karel Lebeda, Jiri Matas, Ondrej Chum (2012)

Key points

LO-RANSAC = RANSAC + local optimization step (iterative least squares)
- Slow in cases of many inliers (either absolute or relative), so we want to improve
LO+-RANSAC:
- Very stable (almost non-random)
- Limiting the number of inliers leads to reduced execution times + better precision
- More robust to threshold choice (due to new cost function)
- Better starting point for bundle adjustment
LO'-RANSAC: lightweight version for "easy" matching (no large motion/viewpoint changes, illumination)
- Up to 6x faster!