In this remarkable demonstration, Daniel Cohen-Or and colleagues demonstrate a method of digitally inferring a three-dimensional solid from a 2D photo. (
Direct link to video). This could be a big timesaver for digital modelers and matte painters.
--------
Thanks, Rob Nonstop
I saw this the other day. Was pretty amazed by it. I also enjoyed that they showed the failures of some of the AI.
ReplyDeleteReally neat technology!
ReplyDeleteI find it distracting that the narrator sounds like a 12-year-old reading from a script. I realize this style of speech is common in software demos nowadays, but I haven't gotten used to it.
Interesting trick. I notice that they use exclusively photos of objects comprised of round or rectangular parts; apparently the inference algorithm doesn't work well for other shapes.
ReplyDeleteIt also cannot infer lighting, reflections or transparency. Note how the bottle gets the image of its bottom seen through its side painted on that side in 3D. Or how the obelisk shadow rotates with it. Or how the copied faucet gets out-of-context lighting on it.
So it strikes me more as a way to edit 2D photos than a way to build 3D models. But still, could be a timesaver for someone who needs to move a bottle in a photo just a bit.
Reminds me of Eron Gjoni's Gifted Apprentice, https://youtu.be/hSmsDwvofVU, posted back in 2015.
ReplyDelete