| |
|
| |
|
Erik Learned Miller
University Talk
Detection, Alignment, and Recognition: Relationships and Synergies ( Joint work with Andras Ferencz,
November 3, 2006 10:00 am
Abstract:
In this talk, I discuss recent developments in our "hyper-feature" system for object identification, and in particular our recent work on face recognition. The goal of hyper-feature recognition is to build a custom model for each face we wish to recognize from just a single example of the face. To achieve this, we must bring detailed knowledge of a larger class of objects (the set of all faces) to bear on the problem. In particular, we must develop a generic model of the distinctiveness and repeatability of facial features that allow us to build an effective model of a face from a single example. Our hyper-feature model requires pre-detection and pre-alignment of faces to a canonical position. The better this alignment, the more accurate our hyper-features will be, and the better our face recognizer performs. We discuss recent work in unsupervised alignment algorithms (congealing) as applied to complex images such as faces in general backgrounds. We give a general procedure for converting ANY model of faces which produces a scalar score into a face alignment algorithm that is fast and robust to local minima. The ideas presented in both parts of the talk are generic and apply to any category of objects which can be roughly registered.
If you have questions, or would like to meet the speaker, please contact Ponda at 4-1994 or pondabarnes@tti-c.org. For information on future TTI-C talks or events, please go to the TTI-C Events page.