The auto-modelling tool seems very interesting. As anyone thought of using some kind of eigenmeshes to find a more realisitic face ?
My idea is the following:
- You consider the n vertices involved in the face as a 3n coordinate vector.
You compute a principal component analysis over the gallery of faces already available
Any new face can then be smoothed by projecting it over the few first eigenvectors of the covariance matrix
Things that could make this method untractable could be a too high number of vertices, a too low number of base faces.
Another possibility would be to use M meshes as base.
let save you have a set of M base meshes. n is the number of points for the mesh. Consider the matrix F (3n x M) whose column are the vectors representing the mesh of a face. You can compute an orthonormal base spanning the same subspace as F for instance by computing its singular value decomposition : F = U*W*V^T . U is then the wanted base.
if f is a new face, it can be smoothed over the known base by f' = U*U^T*f.
The new face f' should be more regular than f since it has been projected on the base on known faces...
Well, that's just a suggestion...
Alexis