CAD-based feature-utility measures for automatic vision programming

Three feature-utility measures-detectability, reliability, and error rate-are defined and related to the execution time of a vision program. Using a realistic vision task as a scenario, the authors show how the utility measures can be used to automatically generate the most efficient vision programs. They also present the experimental result of estimating the feature-utility measures using synthetic images rendered from CAD models. The discrepancy between the estimated utility measures obtained from CAD-rendered images and from real training images suggests that one must be cautious in constructing CAD-based vision systems.<<ETX>>