Carnegie-Mellon creates program that generates 3d image from a single 2d photograph

[ashrafazlan]

cool stuff, can't wait for it to be available to the public

[Gentle Fury]

im dling this thing now...I wanna see if it really works.....there was another app from the company that made retimer that claimed this ability, but in reality all you could do was create basic models by hand and apply the videos texture to it as it moved in 3d space.....not very impressive. Interested to see if this actually creates proper geometry as it claims.

[Ninjas]

I'm sure it does, at least to the extent that they show in the photos, which are all screwed up looking. The building is more or less box shapes, you cant get much easier than that!

Still, I'm sure once this tech evolves it will be cool to watch old reruns of The Love Boat on my 3D tv.

[JMcWilliams]

I'm suprised we haven't seen any silly "we don't need modellers anymore!" statements that usually pop up when something like this appears. :)

[thk]

I'm suprised we haven't seen any silly "we don't need modellers anymore!" statements that usually pop up when something like this appears. :)

We don't need modelers any more! We just assume that everything is made of straight horizontal and vertical lines, the ground is flat, and occluded surfaces don't exist. :D

Actually, this is pretty nifty research. It certainly has its limitations, but it's a nice fresh approach for designing an intelligent vision system.

[rootdown]

My favorite part is that it's a learning software that becomes smarter the more it looks at stuff. You know, like the Terminator. Before too long it'll be stealing cars by finding the keys hidden in the glare visor.

I also dig that one of the pictures they used it on is obviously a painting. Give it a bit more development time and you could do all your simpler blocking-in with a pencil and a scanner.

[yinako]

this works like camera projection, no?

[parallax]

It simply can't be done. ever. The only remote possibility of this working, is using an endless database of reference images.

[thk]

It simply can't be done. ever. The only remote possibility of this working, is using an endless database of reference images.

Well, our brains don't have these endless databases of reference images, but we seem to do all right.

[JMcWilliams]

We don't need modelers any more! We just assume that everything is made of straight horizontal and vertical lines, the ground is flat, and occluded surfaces don't exist. :D

Actually, this is pretty nifty research. It certainly has its limitations, but it's a nice fresh approach for designing an intelligent vision system.

Yeh, actually my first thought was that it would be a good starting point for lifting matt paintings onto geometry. Though of course, CG is really not what they are working on this for, we can adopt it. :D

[innerminds-i]

It simply can't be done. ever. The only remote possibility of this working, is using an endless database of reference images.

yea i dont like the idea either, seems way too fiddly. Youll never be able to do it perfectly. It will always rquire extra work from the user unless the object is very simple.

Id prefer to keep brushing up my modelling skills instead of wasting time learning to use this program.


just my 2 sterling pence

[TheLostVertex]

Well, our brains don't have these endless databases of reference images, but we seem to do all right.

Actually, for the most part our brains do. But more importantly we have stereoscopic vision which is what gives our mind enough information to understand depth. With out 2 eye(or cameras) proper depth perception wouldnt be possible with out extra equipment(range finder, etc). I knew a girl who was blind in one eye, no depth perception at all. I always wondered if she got pissed of when going to something with 3d glasses :D

-Steven

[thk]

Actually, for the most part our brains do. But more importantly we have stereoscopic vision which is what gives our mind enough information to understand depth. With out 2 eye(or cameras) proper depth perception wouldnt be possible with out extra equipment(range finder, etc). I knew a girl who was blind in one eye, no depth perception at all. I always wondered if she got pissed of when going to something with 3d glasses :D

-Steven

Yeah, a one-eyed person wouldn't have much luck w/ 3D glasses.

Given that there are many people w/ one eye or eye patch (arrr! shiver me timbers!) who manage to navigate the world without severe issues, I'd say that stereoscopic vision isn't as critical as spatial intelligence and geometric knowledge. Just do a simple experiment: cover one eye, then look on your desk and reach for various objects. It feels a bit odd, but you can still do it, because you can parse numerous other visual cues, not to mention a priori knowledge of those objects and an expectation of how they're arranged. Similarly, vision research offers algorithms for reconstructing depth from stereo, shading, defocus, templates, projection, and many others. Our brain uses numerous different methods for inferring 3D from 2D, not just stereo reconstruction, which is why you can draw a 3D-looking sphere on a flat piece of paper by shading it just right. Or, for that matter, look at a photograph of a building and imagine what it would look like from different perspectives. And that's exactly what this CMU paper is exploring, using some reasonable assumptions and clever use of standard machine learning techniques, which is why it's cool.

[chadtheartist]

This doesn't look like it does much more than ImageModeler (http://www.realviz.com/products/im/index.php) from Realviz. Other than it being automated of course.

Would be great for adding depth to Matte paintings.

[paintbox]

I think they should actually work on faces, since they are more or less symmetrical...

A building could be anything at the parts that are obscured.

[ThePumpkinKing]

We don't need modelers any more! We just assume that everything is made of straight horizontal and vertical lines, the ground is flat, and occluded surfaces don't exist. :D

Actually, this is pretty nifty research. It certainly has its limitations, but it's a nice fresh approach for designing an intelligent vision system.

Don't forget the fact that the computer will probably also know exactly what kind of topology the person will need in there model. You know, cause it can read minds. Modelings programs and the people who learned to use them are obsolete, long live photo modelling programs and the people who've learned use them.

[PhantomDesign]

It is theoretically possible to do a semi-accurate "reverse render" - converting a still imge into a 3D model. It would be a rather difficult opperation, but once you sucessfully start seperating reflections, shading, and other surface properties & calculate the location of relevant lights . . . it seems like it would be possible. You would have missing geometry in various locations, but that's like asking someone to discover what's on the other side of a wall without extremely specail equipment. I suppose eventaully, we might even be able to extrapolate additional scene elements that are in reflections.

[CGTalk Moderation]

This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.