Abstract for October 11, 2005

Evgeniy Bart (IMA): Object recognition and classification with limited training data

Learning a visual task frequently requires a large training set, which may be costly to obtain. In this talk, we suggest an approach to reducing the required amount of training data. The approach is based on reusing experience with already learned tasks to facilitate learning the novel task. This general method is illustrated on two specific visual tasks.

The first task is object recognition across variations of viewing conditions (such as viewpoint). Experience with familiar objects of a certain class (such as faces or cars) is used to facilitate generalization to previously unseen views of novel objects of the same class. In the resulting scheme, a face that has only been seen in a frontal view is successfully recognized in profile. Pose, illumination, and other viewing conditions are handled in a single general framework.

The second task is object classification. The goal here is to observe a single instance of a novel class, and to generalize to additional instances of this class. Experience with already learned classes is used to facilitate this generalization. Both high-level data (on the level of entire classes) and middle-level data (on the level of individual features) help improve generalization. Combining the two sources of information further improves the performance.

Joint work with Shimon Ullman.

Postdoc Seminar Homepage