Generative Models for 2-D images of 3-D scenes

Generative Models for 2-D images of 3-D scenes Anitha Kannan Doctor of Philosophy Graduate Department of Computer Science University of Toronto 2007 This thesis introduces generative models of appearance for analyzing 2D images of 3D visual scenes. Many 2D images contain multiple objects, where image components corresponding to each object undergo deformations, eg., due to changes in the shape and appearance of the object, non-uniform scaling caused by camera zoom and global transformations such as changes in the location of the object. In this thesis, we use Bayesian networks to model these sources of variability and their potentially noisy interactions while describing an image. Given an observed image, the model is used to infer the distribution over the possible explanations of the unobserved sources. But, there is an exponentially large number of combinations, so an exact search over these combinations is practically infeasible. This thesis introduces approximate inference and learning algorithms based on variational methods for inferring the underlying explanations of the observed data. In the first part of this thesis, a fast unsupervised algorithm for learning an object’s appearance in a linear subspace, while being invariant to global transformations such as translations, rotations, and scalings is introduced. For modelling multiple objects in the scene, the second part of the thesis presents a generative model for explaining an image using a layered composition of “card-board cutouts”. With this, each object is accounted for by a model of 2D image, including a transparency map that specifies pixels belonging to the object. Each such “layer” includes hidden variables that account for sources of variability such as changes in appearance, shape, deformation and global transformation. The model is learned using a