Physically Accurate Synthetic Images for Computer Vision System Design

The design of a computer vision system for part presentation is a complex hardware/software problem. In the past, standard renderings of parts available from CAD systems have been used as aids in the design process. However, such standard renderings are very limited because of the simple illumination models they employ. We present preliminary results of a study of the utility of physically accurate synthetic images in the design of vision systems. Physically accurate images can potentially be used both for the hardware lighting and sensing design, as well as for template design for model-based matching for part location. We describe how state-of-the-art computer graphics global illumination algorithms can be used to generate images for the vision problem. We present a comparison of a variety of synthetic images to images captured using the GRIPPS retroreflective vision system under development at Georgia Tech