fa14d4fe2f19414de3ebd9f63d5c0169-Reviews.html

This paper proposes a general method for solving image-based computer vision tasks using a generative probabilistic model that uses a graphics program to generate images. The method takes the standard Bayesian approach to frame the inference of the target variables, and uses Metropolis-Hastings to perform the inference. This framework is implemented for a CAPTCHA and a road-finding application, with favorable results reported for each one. The primary contribution of this paper is a proposal for using graphics programs as a key element of a generative model for image-based tasks. While their claim that there are no previous real-world image interpretation frameworks that combine computer graphics among the other elements they list (last paragraph of Section 1) seems accurate, their proposed system does not seem to qualify as such a framework unless it's under a restricted interpretation.