In traditional face recognition, a 2D test image of a person is matched against 2D reference images of various persons stored in a database. An inherent problem of this 2D-based task is that the test person usually takes a different position in front of the camera each time it wants to be recognized, and especially different from the position the reference images were recorded in. This talk explores a novel face recognition strategy using a spatial model of the head together with a corresponding texture map to create arbitrary views of the person, and thus assimilating the two poses of the reference and the test image. The main topic of this work is the reconstruction of a head model of an individual based completely on vision techniques, namely the integration of multiple visual shape cues. A framework, in which such an integration could happen, is presented together with an example implementation subsequently integrating shape from contours and shape from luminance, a new voxel-based approach. The resulting models are used for a test comparision of traditional face matching with the model-enhanced approach.
This talk will be held in the Main Lecture Hall at ICSI. 1947 Center Street, Sixth Floor, Berkeley, CA 94704 (On Center between Milvia and Martin Luther King Jr. Way) [http://www.icsi.berkeley.edu/location.html]
If you are interested in checking the availability of papers regarding this talk, please contact the speaker (whose e-mail address appears at the top of this message).