Errm.
I know this is old and forgotten and the original poster will never read this, but this might be of use to anyone with the same question who comes along.
So my understanding is that you do not want to use an existing 3D engine. If you do, read the above posts and stop there; the rest depends on the engine and nothing more.
But for any who want to make this work all from scratch for any reason (perhaps as a learning quest) then you will find daunting but not insurmountable tasks ahead of you.
Firstly, yours (and anyone else's) idea of simply taking a 3D model of a building and generating some kind of data to use to be able to collide with the walls and walk around in it is simply wrong. It will never happen. Sorry for the bad news, but your model is completely useless.
Terrain, static objects (including buildings), and movable characters are distinctly separated in the game world and for a very good reason.
Each one has certain known properties that allow certain types of shortcuts to work on them. These shortcuts are not just good at improving the performance of the game, they are vital.
The shortcuts/methods of dealing with the static data that you have involve a series of tests against closed polyhedra such that the polyhedra are of a known class (sphere, box, etc.) and are known to be convex.
Map editors typically allow you to create blocks, chop parts away or modify them a bit, then either remove that shape from space or add it to it. If you ever create a concave shape the editor will whine.
The map editor stores both the final result in plain polygons and also the history of additions and subtractions. When it comes time to perform hit detection against any part of the building, it runs down the history of the building and performs simple "inside of" and "outside of" tests on the simple shapes you used to create the map.
It becomes clear that in order to do any kind of hit-testing on the walls of the building you need a lot more information than what you have. Your simple model is nothing but image data. There is nothing there you can use to create a stable/decent/more-than-1-frame-per-second hit-test.
L. Spiro