Automatic singularity test for motion analysis by an information criterion

Abstract

The structure-from-motion algorithm from two views fails if the object is a planar surface or the camera motion is a pure rotation. This paper presents a new scheme for automatically detecting these anomalies without using any knowledge about the noise in the images. This judgment does not involve any empirically adjustable thresholds, either. The basic principle of our scheme is to choose a model that has "higher predicting capability" measured by the geometric information criterion (geometric AIC).