Realistic range rendering for object hypothesis verification

Abstract

In many model-based object recognition systems, a synthesize-and-verify technique is used to evaluate the quality of hypotheses. This technique synthesizes images of hypothesized objects in hypothesized poses, and compares them against the input imagery, producing a matching score. In this paper, we examine the image synthesis process in the context of triangulation-based range finding. We motivate the use of synthetically shadowed range data for verification, present a simple algorithm for generation of shadowed range imagery, and demonstrate its usefulness in a set of experiments on real imagery.