That fits a bit. My nephew today told me, he would like a game, where he can draw a racecar (2D) and then be able to drive the car in the game. Sadly the demo is for faces only, otherwise I might have hacked something together to make it work ...
> where he can draw a racecar (2D) and then be able to drive the car in the game
If your nephew just would draw outlined 2D car side view, then its would be easy to convert it to 3D by extruding and in one two clicks export to supported format for game.
Just two apps needed: Inkscape + Blender
P.S.: If your nephew no need to play 3D game and top-view racing game (such as DustRacing2D[0]) is enough, possibly only Inkscape would be needed to draw top view of racingcar for game.
Yeah, I know, but his car designs usually also involve the front (it is all about the gimmicks, fire and razors and I don't know what else).
So a neuronal network would definitely not be able to make a perfect model out of it, but maybe enough to have a somewhat blurry version in tux cart for example.
But I will teach him blender soon, so we will get there eventually, but doing it automated, would be very cool, too ;)
I've seen a system that was more rules based that did this based on a set of sampled head shapes and then you could fine tune the results by hand. One thing that I noticed is that creating a mappable texture from a photo and mapping it onto any reasonable 3D head model was already cool. In any case, this particular attempt doesn't look better than previous techniques I have seen. Every single head in this demo had a pronounced sloping forehead which surely can't be the case for every input.
> we exploit the fact that many object categories have a bilateral symmetry. Assuming an object is perfectly symmetric, one can obtain a virtual second view of it by simply mirroring the image and perform 3D reconstruction using stereo geometry
it says stereo so you need two pictures? I was thinking about the photogramemtry technique, this is great... I mean I wonder how accurate it is(with regard to depth calculation from images on a larger scale like feet) but man impressive
No, that's the brilliant part. They use one image and apply the assumption that it's a picture of a bilaterally-symmetric physical object in order to produce a synthetic second image.
hmm probably have to watch the video more thoroughly it seems like you're estimating/guessing the actual dimensions without a new source/angle to compare for the depth aspect. showing my ignorance here -- trained, I guess it depends on margin of error/does it matter, it seems pretty accurate
anyway this would be great with regard to reducing camera count if you're not relying heavily on LIDAR or "physical time to flight" sources of measurement
What makes this especially disturbing is that pupils are sticking out in almost all the examples.