Feature saliency from noise variations in invariants

Abstract

Object localisation and recognition are fundamental problems in computer vision. The goal is to enable a mobile robot to locate general, non-polyhedral objects in complex settings. This requires considerable robustness and reliability, and so low level invariants are used as a robust starting point. In particular, a set of quantities is developed that are both geometrically and photometrically invariant. They are arranged as the components of a description vector, which are matched to locate model instances.

The paper analyses the variations of the invariant quantities. These arise in practice due to image noise and spatial quantisation: the case of image noise is treated here, quantisation being the subject of ongoing work. The noise models obtained show good agreement with experimental results. A probability model for the variation of the description vectors is derived and used to define a saliency measure in the image. Combining this with a Non-Uniform selection strategy in a modified RANSAC (NU-RANSAC) scheme leads to a dramatic improvement in the probability of correctly matching points, which is the basis of localising the desired object.