Extending RANSAC-based estimators to handle unknown and varying noise level

The robust parameter estimation methods are a general tool in computer vision, widely used for such tasks as multiple view relation estimation and camera calibration. In this paper, a new robust maximum-likelihood estimator AMLESAC is presented. It is a noise-adaptive version of the well-known MLESAC estimator. It adopts the same sampling strategy and seeks the solution to maximize the likelihood rather than some heuristic measure. Unlike MLESAC, it simultaneously estimates all the noise parameters: inlier share $\gamma$, inlier error standard deviation $\sigma$ and outlier probability density $1/\nu$. Effective optimization for the computation speed-up is also introduced. Results are given both for synthetic and real test data for different types of models. The algorithm is demonstrated to surpass the previous approaches for the task of pose estimation and provides results equal or superior to other robust estimators in other tests.