On Evidence Assessment for Model-based Recognition

Evidence assessment for model based recognition is concerned with determining if a set of correspondences between image features and model features gives sufficient evidence for recognition. While many previous studies have addressed strategies to establish the correspondences, postmodel-matching evidence assessment has remained largely primitive and ad hoc. This paper presents a novel two-stage scheme of evidence assessment for model-based vision based on: (i) evidence against coincidental configuration of random image features and (ii) evidence against mis-recognition of other objects. We demonstrate this scheme for model based 3D recognition from 2D image features.