MIT talk: Object Class and Subclass Recognition Using Relational Object Models

Speaker: Aharon Bar-Hillel , Hebrew University
Date: Wednesday, February 15 2006
Host: Prof. Tomaso Poggio, M.I.T., McGovern Institute, BCS & CSAIL

Abstract: In the first part of the talk I will present a new learning method for object class recognition, combining a generative constellation model with a discriminative optimization technique. Specifically we use a 'star'-like Bayesian network model, but learn its parameters using an extended boosting technique which iterates between inference and part learning. Learning complexity is linear in the number of model parts and image features, compared to an exponential learning complexity for similar models in a generative framework. This allows the construction of rich models with many distinctive parts, leading to improved classification accuracy.

In the second part of the talk I will address the problem of sub-ordinate class recognition (like the distinction between cross and sport motorcycles), relying on the above-mentioned learning technique. Our approach to this problem is motivated by observations from cognitive psychology, which identify parts as the defining component of basic level categories, while sub-ordinate categories are more often defined by modified parts. Accordingly, we suggest a two-stage algorithm: First a model of the inclusive class is learned (e.g., motorcycles in general) using the technique introduced earlier, and then subclass classification is made based on the part correspondence implied by the model. The two-stage algorithm typically outperforms a competing one-step algorithm, which builds distinct constellation models for each subclass. This performance advantage critically relies on modeling of the spatial relations between parts, and on having models with a large number of parts.

The talk is based on a joint work with Tomer Hertz and Prof. Daphna Weinshall.