If you use your fake-method with terrain which is inwardly pulled against a plane, you can only achieve good results if you use a very high detailed terrain. The same for models. This could be quiet nasty, especially because the areas where "the floor will be" are not orthogonal at all.

When going the pixel way, you have a high degree of details. The collision detection is a good point, but doable.

Have you already achieved some promising results?