Estimating facial pose from a sparse representation
A method for accurately estimating a pose of the human head in natural scenes utilizing positions of the prominent facial features relative to the position of the head. A high-dimensional, randomly sparse representation of a human face, using a simplified facial feature model transforms a raw face image into sets of vectors representing the fits of the face to a random, sparse set of model configurations. The transformation collects salient features of the face image which are useful to estimate the pose, while suppressing irrelevant variations of face appearance. The relation between the sparse representation of the pose is learned using Support Vector Regression (SVR). The sparse representation, combined with the SVR learning is then used to estimate a pose of facial images.
1. A method of estimating a pose of a human head in natural scenes comprising the steps of:
generating, a sparse representation of a human face by transforming a raw facial image into sets of vectors representing fits of the face comprising fits of whole facial features that represent the geometry (position, size and orientation) of the features, to a random, sparse set of model configurations; wherein the sparse representation is a collection of projections to a number of randomly generated possible configurations of the human face, wherein irrelevant variations of face appearance are suppressed by the application of filters whose shapes are matched to that of facial features and the filters are generated to match the variety of actual facial feature shapes of individuals after an out-of-plane rotation.
training, the sparse representation to a set of face(s) in known poses; and
determining, a pose of a head by comparing the trained representation(s) to a facial image.
2. The method according to claim 1 wherein the transforming step further comprises the step of:
collecting, salient features of the face image which are useful to estimate the pose of the face.
3. The method according to claim 2 wherein the training step further comprises the steps of:
learning, using Support Vector Regression (SVR), a relation between the sparse representation and the pose(s).