|
Paul Viola
MERL
More Features Than Pixels: A Framework for Constructing Fast and
Robust Visual Recognition Algorithms
All learning approaches require the acquisition of positive and
negative examples, the selection of a suitable feature set, and
selection of a learning algorithm. I will demonstrate that a wide
range of visual recognition problems can be solved by adopting a huge
and varied set of visual features. A variant of AdaBoost is then used
to select a small set of features critical for a given task. The
AdaBoost procedure provides theoretical bounds on training error,
generalization error, and the number of required features. Other
theoretical results ensure that the presence of spurious features will
not reduce testing error significantly.
We have experimented with as many as 1,000,000,000 binary features on
problems with thousands of examples. Though the learning process can
take days of computation, the resulting classifiers are extremely
efficient and robust. Using this approach we have developed the
world's fastest face detection system, which can find faces at 15
frames per second frames on a laptop. We have also applied similar
ideas to problems in image database retrieval. New results include a
very fast system for facial analysis, which can locate the eyes and
nose, and can guess the gender of the user. I will demonstrate these
systems during the talk.
|
