Uncertainty reasoning in object recognition by image processing

Abstract

Object recognition in digital images is a primary issue in robotics. We consider the model-based vision problem, where objects to be recognized come from a database of geometrically precise models. However, the modeling process involves uncertainties, and thus predicted collections of features will be subject to possible variations. Likewise, the image analysis problem using digital images must deal with sensor noise and ambiguity in the imaging process. Accordingly, object recognition is not a simple matter of matching features sets, but must deal with variabilities in the models and in the extracted features in the scenes. In this paper, we consider how these uncertainties should be handled. We describe how predicted variability can be used to compute a match metric, in order to assess the quality of possible models. We discuss two methods for dealing with extracted uncertainty. Finally, we speculate on other methods of assessing uncertainty in the recognition process.

Discussions with Davi Geiger and Martin Garcia-Keller are gratefully acknowledged. The author can be contacted at hummel@cs.nyu.edu.