Tracking subspace representations of face images

Abstract

A small collection of successive frames of a video sequence of a talking person defines a subspace whose dimensionality is much less than the size of one frame. Any face image drawn from the video sequence can be associated with a subspace defined by itself and the frames close to it in time. Representing faces in their associated subspaces can reduce the complexity of further processing that is done on the faces. For a slowly varying video sequence, the subspace that a sliding window of frames will define will also be slowly varying. We describe one technique for efficiently tracking slowly varying subspaces. The method builds and updates the subspace of face images using an adaptation of the Gram-Schmidt orthogonalization procedure. We outline techniques for reducing the dimensions of the subspace while keeping the MSE small, and present one measure for dealing with the perceptual quality of the reconstructions. Finally, we consider subspace tracking in the context of video compression. >