UCLA

Scene understanding is a fundamental problem in computer vision research. We

address this problem in an “analysis by synthesis” fashion - explain observed data

(an 2D image) according to a set of spatial grammar (describes the underlying

functional arrangement and 3D geometric structure of a scene) that generate it.

The inference process is carried out in a Bayesian framework. The posterior

probability includes a prior probability reflecting the knowledge of indoor 3D scene

structure encoded by grammar, and a likelihood that evaluates the accuracy of the

re-projected image and the physical plausibility. The most reasonable explanation

of the image is given by a parse tree that maximizes the posterior probability, and

it is found by reversible-jump Markov Chain Monte Carlo sampling.