Towards a Vision Algorithm Compiler for Recognition of Partially Occluded 3-D Objects

Our goal is to develop a model-based vision system that is capable of recognizing 3-D objects in range images in spite of partial occlusion of the objects. We present new methods for object recognition and localization and describe the implementation and performance of these methods in our Vision Algorithm Compiler VAC model-based vision system. The VAC is given a sensor model and a set of geometric object models and generates a recognitionlocalization program for the specified objects in images from the specified sensor. Our recognition algorithm is based on the hypothesize-and-verify paradigm. We use the sensor-modeling approach to build accurate models of our prior-knowledge constraints that account for constraints due to sensor characteristics feature-extraction algorithm behavior, model geometry, and the effects of partial occlusion. We phrase the hypothesis-generation process as a search for the most likely set of hypotheses based on our prior knowledge-in contrast to typical constrained combinatorial searches.