This paper is from INRIA, the French research institute, and presents a new algorithm for automatically constructed low-poly standin models. In particular, they started from the idea of billboard standins, as games routinely use for distant objects.

However, instead of using a completely fixed geometry (like a single billboard or a crossed pair of billboards), they compute a smallish set of billboards which can well approximate the model, then render new alpha textures for that set of polys. The result is a greatly reduced model that reproduces the original -- in particular, one of the impressive things about this technique is how well they reproduce the silhouette of the original. Their examples included a helicopter reduced to 32 polys, a dinosaur reduced to 110 polys, and a very impressive 40-poly Eiffel Tower.

To select the billboards to be used, they have a greedy optimizer algorithm that runs in plane space -- it computes a coverage metric and a penalty metric for input polys that are "just missed", then repeatedley selects the best output billboard and removes those input polys from consideration. The only input parameter is the maximum allowable error in object space for vertex position -- you don't explicitly set the number of output billboards you want, but you can tweak the allowable error parameter to get the size you want.

One of the biggest advantages of this algorithm is that it doesn't require closed meshes; it operates just fine on edges with boundaries, because in effect the simplified meshes can move boundary edges, which none of the mesh simplification algorithms can do. That makes it seem a lot more applicable to the kinds of models we build, which are rarely one closed manifold. It would also be possible to run this algorithm per subset and actually get articulated standins (because it doesn't care about non-closed meshes).

The biggest disadvantage of this algorithm is that like other billboard schemes, because new textures are rendered for each output billboard, it does not re-use the textures of the original model, and eats up new texture memory instead. However, that's probably not so bad when used to generate LODs, since the algorithm constructs textures that are uniformly dense in object space. When building LODs, the required texel density will presumably be very low, and thus the size of the new textures will likely be very modest.

The full title of this paper was "Billboard Clouds for Extreme Model Simplification" by Decoret, Durand, Sillion, and Dorsey.