Computer vision systems use image features to detect and categorize objects in visual scenes. Adequately capturing the rich variability of natural environments is challenging in part due to the scarcity of carefully labeled training data. My research has explored hierarchical models which use contextual and geometric relationships for more effective learning from large, partially labeled image databases. By leveraging nonparametric Bayesian statistical methods, which generalize the Dirichlet process, we can efficiently learn models whose complexity grows as more data is observed.
In this talk, I begin by describing hierarchical models relating objects, the parts composing them, and the scenes surrounding them. By coupling part-based models with spatial transformations, the transformed Dirichlet process shares knowledge among related object categories, improving categorization performance. I then show how related ideas can be effectively used for three-dimensional geometric reconstruction, low-level image analysis and denoising, and segmentation of complex scenes. These projects integrate variational and Monte Carlo methods in novel ways to develop computationally efficient, scalable learning algorithms.