Adaptive estimation of error sigma using cumulative density ratio test in RANSAC

Abstract

One of the most important issues in RANSAC is to estimate error threshold robustly. In this work, we propose an adaptive method for estimating error distribution in RANSAC using a cumulative density ratio test. In the proposed method, we first compute cumulative histogram of residual errors of data and then estimate error distribution by maximizing a similarity measure between the computed cumulative distribution and the standard normal cumulative distribution. The similarity between two cumulative distributions is measured by a cumulative density ratio test. The experimental result confirms the effectiveness of the proposed method.