Machine Learning and Friends Lunch

Learning Object Appearance Models via Transformed Dirichlet Processes

Abstract

Object recognition systems use features extracted from images to localize and categorize objects. Such systems must be robust to the rich variability of natural scenes, and the often small size of training databases. In this talk, we describe a family of hierarchical generative models for objects, the parts composing them, and the scenes surrounding them. We employ Dirichlet processes to learn flexible appearance models which transfer knowledge among related object categories. By coupling these part-based appearance models with spatial transformations, we also consistently account for geometric constraints. Using Monte Carlo methods, we then adapt these transformed Dirichlet processes to categorize objects given few examples, and automatically recognize groups of objects in complex visual scenes.

Joint work with Antonio Torralba, William Freeman, & Alan Willsky at MIT.

Back to ML Lunch home