Evolution Strategy Sampling Consensus for Robust Estimator

Abstract

RANdom SAmple Consensus (RANSAC) has been applied to many 3D image processing problems such as homography matrix estimation problems and shape detection from 3D point clouds, and is one of the most popular robust estimator methods. However, RANSAC has a problem related to the trade-off between computational cost and stability of search because RANSAC is based on random sampling. Genetic Algorithm SAmple Consensus (GASAC) based on a population-based multi-point search was proposed in order to improve RANSAC. GASAC can improve the performance of search. However, it is sometimes difficult to maintain the genetic diversity in the search if the large size of outliers is included in a data set. Furthermore, a computational time of GASAC sometimes is slower than that of RANSAC because of calculation of the genetic operators. This paper proposes Evolution Strategy SAmple Consensus (ESSAC) as a new robust estimator. ESSAC is based on Evolution Strategy in order to maintain the genetic diversity. In ESSAC, we apply two heuristic searches to ESSAC. One is a search range control, the other is adaptive/self-adaptive mutation. By applying these heuristic searches, the trade-off between computational speed and search stability can be improved. Finally, this paper shows several experimental results in order to evaluate the effectiveness of the proposed method.