Automation: 2D to 3D

http://www.youtube.com/watch?v=VuoljANz4EA

This is really cool. I wish I could find the code they used to do this.

I'm pretty sure it's a matter of them stretching the image... something about the 3d simulations didn't look right

I assumed they were analyzing the shading. Thats the only way I can see it working out without specifying whats an object and what is't. But even then, assoiating one object to another that ight be way far ahead of the one behind it would take something else.

They aren't actually making those objects properly 3D. As timeshifter said, they're just being stretched and then mapping the bitmaps onto a pseudo-3D surface.

What it looks like they've done is some sort of object recognition to see where the object is and what the floor is like in relation to that. Then generated a simple 3D image behind that and mapped the parts of the bitmap onto it.

The train and the first street scene are basically the same - there's an object strecthing along the LHS of the picture, so make that the "wall", make the end of the train (or the side street) the "back" and the floor is as is. Then just move your camera position and redraw the bitmap.

The second street is similar but with objects along both sides.

The ship and the building are both objects in the middle with space down both sides.

Nevertheless, it's a good start.

The hardest part would probably be identifying the vanishing point(s) of the scene. Once you have that ten you can use edge detection to find convenient rectangula areas. Its a nice effect but not impressive enough in its current form.

Hear hear.

Rhubarb.

Guacamole