One of the popular approaches to inverse rendering of a face surface by a photo
portrait is synthesis-by-analysis. The method is to render existent 3D face
models under different conditions (different parameters of illumination,
rotation, position etc.) to find those which produce an output that is the most
alike to the given input. In order to increase robustness of the method, it's
usual to find facial landmarks to rely on them during the fitting procedure. A
question occurs: what feature points to use? Should we use a tip of a nose,
corners of eyes and lips? What will we lose if we use only a tip of a nose and
corners of lips? The current talk covers the following situation: - we have a
linear generative model of a face surface; - affine transformation is applied
to the face surface; - we can match all needed landmarks ideally; - an input
image is made by orthogonal projection (instead of perspective projection); -
we use only projections of facial feature points (no color information) for
reconstruction. The question is: what points should we choose to find affine
transformation applied to the face and its shape?