^ It's the metric that didn't really make any sense; lwilliams3 said it better by putting next and current gen polycounts and expectations side-by-side but, not only do you have various means to an end, there's also the issue of a 1000 polygon ball is already round enough, so 2000 polygons is bound to be diminishing returns regardless of wether you do it or not; if you don't nobody will notice, and if you do perhaps you'd be better off investing that budget elsewhere, it's that simple; keeping the framerate will be a better balance than running it into the ground because you want to push a few more polygons (you needed to do that on PSone/Saturn and N64 as a few less polygons made such a huge difference, not so these days).
Simpler polygonal graphical styles are technically less in need of polygons to pull through, as evidenced by Wind Waker HD being re-issued and being praised everywhere, that's a game where the Link character model weighted in at 2800 polygons, at least on the very similar GC version.
You could pull a shitload more polygons into it and it would be diminishing gains considering it already looks pretty damn good.
Mario games are a similar example, Mario Galaxy's Mario character model had
4877 polygons, fun fact: that's 3.6 times over Mario Sunshine's 1348 polygons; which could be considered it's "last gen" counterpart.
In order to quadruple that you'd only have to reach 19508 polygons, something they probably only didn't go for because a) they didn't want to b) they didn't need to. 19.5k polygons being above average for current gen standards, specially if it's an in-game character model and not a cutscene one, in fact at 720p more than being a challenge it almost seems like a waste.
Models are pretty high polygon for what they are though, just like Mario Galaxy was pretty high polygon for a Wii game.
Doesn't really make sense to go the "generation route" about this, because it either looks pretty good or it doesn't; if it doesn't and it's not due to lousy modeling/animation/texturing then it boils down to lacking in geometric detail (something that can be attributed to lacking polygons, or being poorly mapped with more advanced techniques, but we already went there).
In the end, it's like this, you have a cube, is a 10.000 polygon cube better than a 12 polygon cube? next gen cubes are still cubes, they don't need to have higher geometry to them (and you've had games built around cubes, like
Cubivore, hell, there's 3d dot games heroes on PS3), so every game has to find their own sweet spot; whatever the polycount Mario 3D world went with... It's probably appropriate as long as you don't feel it's lacking geometry. This generation was 720p, the next one will be sub-1080p-and-often-720p, it's not such a big leap that current generation character models are bound to look like shit providing they were properly detailed and textured, so investing somewhere else to make a difference would be the clever thing to do, unless you really require those extra polygons (I mentioned Forza 5 before, so: Forza 5 seems to need them and hence it'll go there); everybody else... perhaps not, unless on cases where they had variable LOD going on where they'll of course try to stick with the highest LOD from the get go (or apply tesselation, but that might be a game changer in itself).
What defines a generation is not a linear thing; polygons defined it at one point, we're past that point though, and this generation is gonna be all about RAM, shader capability and passes; not polygons.
Backtracking... the "in between PS2 and PS3" remark didn't make much sense. There's lots of ways to claim something and that was simply a bad way to do it; because these days you'll only push for higher *something* if you think you need to, no title this gen tried to be a "polygon pusher" for instance, because even if they hit a wall and have to apply some trickery... they simply won't disregard a balance of needed features, they can't go one trick pony on it in the name of getting horrible results.
Resources are still finite, but precisely because of that you have to pick your battles it just happens that, unlike a few generations ago polygons are often not your biggest problem, far from it, same as saying variable LOD is not a industry standard. (for open-world games though, it is); more power, had that been "the" bottleneck will just enable developers to avoid the variable LOD/proximity asset replacement process and it'll result in generally the very same results to the viewer. Hence, not much of a generational leap, right?
Unless Tesselation comes into full play, but that's a potential game changer for a reason, if used in a plentiful manner it could enable a generational leap on a machine no more powerful than X360, precisely because it can add detail to details close to the camera and simplify detail for further away objects. Nobody knows how to use it in a way that is not complementary as of now though; so games that use it are using it as a icing on top rather than building the game to be light from the ground-up yet really scaleable.
The needs of developers are never really predictable, so before a generation is in full swing we'll never know what kind of advantage really gives developers the edge, but the sheer graphical leap of previous generations is unattainable, and it would be even if the leap could be rated to be as big (in processing power) as in the past something that it really isn't. So they have to focus on the things that really make a difference; such things might not even have to do with power in a linear way, as seen with animations on, say... GTA5; it's like the discussion of going realtime shadow map for a gameworld or keeping it pre-baked, really. They'll only go realtime if the gameworld needs it or they have so much power they don't need it elsewhere - polygons is bound to a similarly etched balance.
And Mario 3D World uses as many polygons as it needs to.