In this paper we consider the full cycle of Image Understanding (IU): the generation of 3D object hypotheses (inverse model) from images and their projections back onto image data (forward model) in terms of Bayesian inference processes. Each subprocess is framed as a local optimization problem based on a component model and observations. The end result is an IU system that not only provides a symbolic description of scenes but also generates fully 3D versions of the scene being sensed, providing validation criteria for image annotation.
展开▼
