Object Recognition with Many Local Features
by Scott Helmer
There has been a great deal of attention focused on part-based approaches to
object classification in recent research in computer vision, and some approaches
have achieved a surprising amount of success. However, learning models with a
large number of parts has been a particular challenge. One of the most
successful approaches is that of \ferg\; \cite{fergus03} who have developed a
generative model for recognition that achieves excellent results on a variety of
datasets. The learning method that they present to learn the parameters for the
model, however, requires an exponential amount of time to train as the number of
parts increase.
In the talk we present an extension of their generative model, and the
development of a learning algorithm that can learn a large number of parts in a
reasonable amount of time. In particular, we have developed an incremental
learning algorithm where the model is initialized intelligently with a small
number of parts, and parts are added to the model one at a time. By taking such
an approach we are able to learn models with a large number of parts in nearly a
linear amount of time in the number of parts. The approach is validated on a
number of datasets, including cars, motorbikes, and faces, and demonstrates
excellent recognition results along with large models learned in a reasonable
amount of time.