Neural network makes 3D images from still pictures

boing
1

Originally published at: https://boingboing.net/2020/03/06/neural-network-makes-3d-images.html

…

2

Honestly, the effect looks dreadful. Everything is warping and stretching in a really weird way and in lots of different directions. Things it tried to move forward got these weird stretched texture lines, and generally looked like they were punched in rather than out.

14 Likes
3

This feels more like a short and unpleasant trip on mushrooms, then anything else.

8 Likes
4

Those are technically pretty amazing, but that bride’s shoulders will haunt my nightmares FOREVER, thanks

6 Likes
5

Déjà vu:

3 Likes
6

Wow, ow, my brain hurts looking at that.

2 Likes
7

The lampposts reminded me of early Fleischer cartoons

5 Likes
8

I think I’ve seen this used for filler on one of the big cable channels, hulu or netflix or some such, with stills from featured shows.

1 Like
9

let me fix that for you:

Neural network tries to make 3D images from still pictures

6 Likes
10

You’d probably hate LSD

1 Like
11

This AI is unwell.
Bender Smash

5 Likes
12

Woooooooosy

13
1 Like
14

The trick is to estimate each point’s distance from the camera, and then interpolate the parallax effects. You can guesstimate the distances by converting blurriness to distance, but this falls apart when things in the foreground are smooth and things in the background are richly textured. It would also fall apart if something is way closer to the lens than it should be. I hate to see what would happen if somebody accidentally put their thumb over the lens and then ran this algorithm.

I suppose interpolating from the different angles is doable if you have something like a convolutional neural network, or a capsule network, but you would first need shit tons of data.

That archway in the first image looked like it was going through itself.

If this is something the researchers did to show that their neural net works, maybe they should stop showing that and use a different image instead. I mean, sure, it took a lot of work to get to this point, but to even the most casual observer that looks fucked up. If your goal is to make a 3D image from a 2D image, they’ve succeeded, but if it’s to make a 3D image from a 2D image and not fuck it up, they failed.

Must have been from Ken Burns’ documentary about hallucinogens.

I noped out at the lampposts. Couldn’t take any more.

15

…These boxes aren’t rotating

16

The title here really ought to read “Neural network makes terrible, disturbing 3D images from still pictures”

17

This topic was automatically closed after 5 days. New replies are no longer allowed.