3-Sweep: Extracting 3D models from (just about) any photo
31 replies, posted
Garry just tweeted this, so I hope it hasn't been posted already. But the video blew my mind and I figured people would like to see it.
[media]http://www.youtube.com/watch?v=Oie1ZXWceqM&feature=youtu.be[/media]
I never imagined this was possible.
FUTURE
There was a program that did this a few years back, but it was really shitty and didn't work most of the time. I doubt this one is all that much better.
This looks way too good to be true.
Are you freaking kidding that is awesome.
This is interesting from an academic point of view, and can help further developments on this area which is great. It could be useful in the future to get a quick rough model but it's still early and unrefined enough for practical uses.
[QUOTE=Edvinas;42131854]This looks way too good to be true.[/QUOTE]
If you google the names of the people behind it, they're all like professors.
And then millions of 3d models of penises were made.
The modeling seems plausible enough, but what really puzzles me are the textures - static images tend to have defined lightsources which are often reflected onto the objects themselves, but the textures this thing applies onto the objects are almost perfect, in that they are stripped of any shadows and are virtually seamless. How the hell did they do that?
Also, I would love to know how they handle stuff like UV coordinates of objects like this on-the-fly.
[QUOTE=Polonium9;42132047]And then millions of 3d models of penises were made.[/QUOTE]
They said that it can make 3d models of [I]man-made objects[/I]. I think it would have issues attempting to model organic objects.
The way it textures it is awesome.
[QUOTE=Rellow;42132136]They said that it can make 3d models of [I]man-made objects[/I]. I think it would have issues attempting to model organic objects.[/QUOTE]
Obviously it is heavily relying on geometry, so natural stuff that just is random on every side won't play well with it.
Pretty sure this is magic.
[editline]9th September 2013[/editline]
In all seriousness though, this is amazing.
This is fucking great
I really want to use this
Now combine these in some kind of database along with tags, like typical colours, shapes, scenes it could be found in etc. Then make a system that will go through every picture on the internet and model everything + get tags of what it is. Then you can snap a picture, and have the program re-create and understand the entire scene in 3D.
Then give robots access to the said database..
GG Humans.
now you don't even have to learn to model, FPS Banana will boom
This only models shapes with radial symmetry, and the way it applies textures is extremely limited, notice how they hardly rotate an object to see the side or back of it, and when they do it simply has mirrored the texture.
Very cool tech, and I'm excited to see where it goes, but it's extremely limited beyond the use of the kinds of simple (and somewhat pointless) edits they made in the video. All the models they made would take about as long to model and texture in a traditional package with minimal training.
I'm going to make a 3D model of my penis and send it to tje women at my workplace
[QUOTE=Jocke;42132654]I'm going to make a 3D model of my penis and send it to tje women at my workplace[/QUOTE]
Hope you get fired.
Since this seems to heavily rely on shadows and highlights, I would like to see this adjusted for paper drawings that you could shadow appropriately for the tech to pickup and then turn your drawing into a 3D object.
thankfully this isn't so good that we're all out of the job but it does mean that the time it takes to make smaller environment props like bottles and streetlights n such is reduced dramatically
Pure fucking magic.
A great step forward!
I don't know what it is called but I remember seeing something similar way back but with video. They walked around a car and placed points in a program, thus building a mesh. Not as fancy as this though.
Maxofs2d modelling secrets revealed.
I'd say give it another 5 years or so, if they keep progressing the technology, and this could become an industry standard. I'm more excited for the future of the tech than what it does now.
I love how they showed its weaknesses, too. It's pretty awesome regardless, you just need a relatively high contrast photo of the object and you've got a model of it.
This is ridiculously impressive.
I'm hoping that they advance it to where you can use multiple photos for better texture scanning instead of just a repeated texture on the other side.
If so, you could essentially model an entire "baked" room based on 3+ photographs. That's fucking awesome.
too bad it doesn't seem to resolve to a model w/ quads, only tris
how do i obtain this software
Sorry, you need to Log In to post a reply to this thread.