Got me wondering how hard that kind of thing is actually. If not for texture tiling, it would be a pretty easy task to take merged objects (split based on location in the track) and compile a list of the materials+textures each one uses. Then you'd figure out a set of atlases that hits some compromise between # of texture atlases and # of materials per dof, assign the textures to locations in the atlases, and rewrite UV coordinates into the texture atlases, while merging materials. Cleaning up the edges so that mipmaps don't go bad would be another concern, either atlas similar textures (trees aren't a problem cause the edges are transparent) or inset each texture with some tiled repetition (like make the actual texture space 200x200, plus repeated around the borders) if necessary.As for the atlas UV'ing part itself, hmmm. I use 3DS Max 2010 here but I know 2011/2012 came with some better UV tools for this exact kind of job... so even for me it's harder work than otherwise.
I know in games like PGR2 for example, they just made everything like we would usually, with small textures and many small items and stuff, and then a big program would optimise/batch/atlas everything and adjust UVs as appropriate.
We don't have that luxury, but in practice it could be an automated process. Even tiled textures can be atlased because our scripts can chop in edges to give UVs something to 'adhere' to as it were.
The trees, vehicles, etc. on this track are prob. an easy candidate for this kind of thing, since they don't use tiled textures. Could, in a first iteration, just check that the UV coordinates start on [0,1] for all polies using a particular material before allowing it to be put in an atlas.
It does add a kink having to manage different materials since that information needs to be fed in somehow. I suppose the dofs could be split to use a single shader in the first place (all solid matte, all specular, all transparencies, all decals, etc.) which from the sounds of it is not much of a performance hit compared to a single dof with multiple materials.
Before tackling the problem of cutting up too-large polygons so that tiled textures can be atlased I'd like to see more evidence that there's a meaningful performance advantage to be gained there. To some extent (maybe up to 2-3x the original texture's size) it might even make sense to, say you have a 256x256 original texture, tile that in the atlas so you have a 512x512 or larger area to work with, and map UV coordinates into that space instead of a single tile. Or if it's something like a road/sidewalk that only tiles in one axis, just run it as an entire row of the texture atlas. As long as you reassigned UV carefully, you could even use the extra space to add variation in the texture.