if you need per pixel depth values, then you have no chance to get what you want from just one picture, there are methods based on stereo, focus, additional lightsources, moire effects, etc. for that, but they all need 2+ pictures.
In some cases you can get Hardware cameras that do it for you (i dont know how these cameras work but they give you depthvalues).
Additionally to this pure optical methods you can get laser scanners, radars, MRIs etc. which use echo-time-measurements (if you know how a radar works, you know how a laser scanner works, the difference is that a radar/MRI can look through objects). Dunno about MRI exactly, but it could work a bit different like SAR or something like that...

In case of faces you can use a different approach. That is you have already a face model and you try to get it mapped onto the face in the image. That can be done by finding some relevant points like eye,nose,mouth,ear positions on the 2d image and then mapping your 3d model onto these points so that they 'fit best'/have a minimal energy. The idea behind that is to 'render' your 3d model into 2d and compare the point positions, then minimize the distance between them, by modifiyng the rotation/translation/scale of the 3d face, and render again until you are satisfied with the result (automatically that means minimum is less than a specified value or hasnt changed since the last iteration). This algorithm can cause some problems depending on your minimizing mechanism. For example you could try to find the next best minimum and it can happen that this 'local' minimum is suboptimal...
finding eyes and other parts of the face could be a bit tricky, but ive seen it working, not only in movies where people like to think of computers as the answer to everything, and randomly clicking on the keyboard makes impossible things possible :)


T2k