It's similar to photogrammetry, but often only a single image is used. You'll take a picture, from a camera/matte painting/whatever, and then project that image through your 3d camera onto untextured geometry in your scene. The idea is to create geometry that is essentially a sillhouette(s) of all the different elements in the scene. Unless there are reflective/transparent surfaces, you won't need anything more than just your image loaded in the texture/diffuse channel, as all your highlights/bumps/etc are essentially already baked in.
While you don't have as much detail as you will if you model everything traditionally, you can still achieve good results as long as you don't swing the camera around too much.
Just search for camera mapping on Youtube for some examples.