Looking for a technology video...

[Capitalizt]

I know this is a shot in the dark, but I forgot where I originally saw this technology and hope someone knows what I'm talking about.

Basically, some company up with an incredible program that incorporates thousands of submitted photographs into a database, and uses those 2D images to create an interactive 3D world where you can zoom in and out of the different locations, changing angles at will.

The program using the best photographs depending on what angle you are viewing...and is able to warp and combine them to create a realistic 3D simulation of those places..

It is mindblowing technology, but I only saw in passing a few months ago..

Does anyone know what I'm talking about?

 

[QuixoticOne]

There are several different implementations of similar ideas, so I'm not sure which one you're talking about.

Most of the space probes have returned 2D image tiles that have then been reprocessed to create a 3D map of the moon / planet / comet / asteroid et. al. for instance.

Here's one that was more popularized in the press recently:
http://research.microsoft.com/...ayArticle.aspx?id=1514
http://labs.live.com/photosynth/

 

[Capitalizt]

thank you...the second one is exactly what I was looking for. I can't wait until we start seeing this level of detail on google maps

 

[BrownTown]

Originally posted by: Capitalizt
thank you...the second one is exactly what I was looking for. I can't wait until we start seeing this level of detail on google maps

you'll be waiting a long time then

 

[NanoStuff]

There was an impressive video that showed not just intelligent mapping of pictures, but showed actual 3D models being generated of the entire scene. This IMO is far more impressive than just ordered scrolling through pictures. You could generate a model of an entire city just by walking through the environment you want to model. Spending weeks in 3D modeling software for a few nice meshes? Forget about it. Will be ridiculously powerful for 3D games when hardware is up for handling such geometric complexity. Flight simulator with the world replicated exactly to the triangle, that would be nice flying.

 

[BrownTown]

Originally posted by: NanoStuff
There was an impressive video that showed not just intelligent mapping of pictures, but showed actual 3D models being generated of the entire scene. This IMO is far more impressive than just ordered scrolling through pictures. You could generate a model of an entire city just by walking through the environment you want to model. Spending weeks in 3D modeling software for a few nice meshes? Forget about it. Will be ridiculously powerful for 3D games when hardware is up for handling such geometric complexity. Flight simulator with the world replicated exactly to the triangle, that would be nice flying.

using pictures is inherently very inaccurate though, you would need hundreds to generate any sort of decent 3D model, it would be alot more effective to use something like laser range finders than a visual picture if you wanted to generate 3D models.

 

[NanoStuff]

Originally posted by: BrownTown

Originally posted by: NanoStuff
There was an impressive video that showed not just intelligent mapping of pictures, but showed actual 3D models being generated of the entire scene. This IMO is far more impressive than just ordered scrolling through pictures. You could generate a model of an entire city just by walking through the environment you want to model. Spending weeks in 3D modeling software for a few nice meshes? Forget about it. Will be ridiculously powerful for 3D games when hardware is up for handling such geometric complexity. Flight simulator with the world replicated exactly to the triangle, that would be nice flying.

using pictures is inherently very inaccurate though, you would need hundreds to generate any sort of decent 3D model, it would be alot more effective to use something like laser range finders than a visual picture if you wanted to generate 3D models.

True, lasers are used today to transform an object into a model, but I can't imagine how fast the scanning rate on that thing would have to be to capture an entire city during a fly-over.

Interpreting that information from multiple images seems much more time efficient. The obvious limitations to mesh resolution with a limited set of pictures would be no issue to games such as flight simulator, 10mm resolution on such a scale wouldn't happen even if the data was available. For games that are more intimate with the ground environment, it's just a matter of getting up close. Seems like a very solid approach with few apparent practical limits considering the availability of images online.

 

[Capitalizt]

Originally posted by: NanoStuff
There was an impressive video that showed not just intelligent mapping of pictures, but showed actual 3D models being generated of the entire scene. This IMO is far more impressive than just ordered scrolling through pictures. You could generate a model of an entire city just by walking through the environment you want to model. Spending weeks in 3D modeling software for a few nice meshes? Forget about it. Will be ridiculously powerful for 3D games when hardware is up for handling such geometric complexity. Flight simulator with the world replicated exactly to the triangle, that would be nice flying.

now that I think about it...you are right. The video I saw was similar to the microsoft demo, only there was no "fuzz" around the pictures being displayed. There were no blank spaces around the images. It was a complete virtual city you could explore...with the different pictures being stretched and blended together as you changed perspective.

It would be incredible if they made this an open-source effort...inviting people to send in their personal photos of different cities and landmarks...and having thousands of them combined into one huge project..

 

[vanvock]

Check this site out, http://make3d.stanford.edu/

 

[NanoStuff]

Originally posted by: vanvock
Check this site out, http://make3d.stanford.edu/

That just folds the image around contrast. Interesting, but mostly useless.

 

[CycloWizard]

Originally posted by: Capitalizt

Originally posted by: NanoStuff
There was an impressive video that showed not just intelligent mapping of pictures, but showed actual 3D models being generated of the entire scene. This IMO is far more impressive than just ordered scrolling through pictures. You could generate a model of an entire city just by walking through the environment you want to model. Spending weeks in 3D modeling software for a few nice meshes? Forget about it. Will be ridiculously powerful for 3D games when hardware is up for handling such geometric complexity. Flight simulator with the world replicated exactly to the triangle, that would be nice flying.

now that I think about it...you are right. The video I saw was similar to the microsoft demo, only there was no "fuzz" around the pictures being displayed. There were no blank spaces around the images. It was a complete virtual city you could explore...with the different pictures being stretched and blended together as you changed perspective.

It would be incredible if they made this an open-source effort...inviting people to send in their personal photos of different cities and landmarks...and having thousands of them combined into one huge project..

I think the one you guys are talking about was an Intel demo from a few years back. I don't remember what it was called or even when it was first released, but that's what's ringing a bell for me.

As for the 3-d construction itself, the efficiency/efficacy will depend on how the pictures are taken. If I take them at regular angles and intervals, it would be relatively straightforward to recreate the object in 3-d.

 

[gorobei]

the general technology is called photogrammetry. Its typically used in 3d film/tv vfx to either matchmove[a 3d virtual camera to duplicate the motion path of a real camera from the real footage] or matchmodel[create a 3d model of an existing envirionment]

there are a couple of commercial software apps that allow you to make basic(flat planes, cubes, and other primitives) models from a collection of images. You basically define track points in each image that are visible in the other images. As long as its a flat wall, the algorithm can extrapolate a position for each point/primitive.

Laser is ok but it generally requires special software to align each scan pass and the coverage always has hole/missing patches. the range and resolutions only really work in medium scale environments(interior rooms and outdoor streets.) and any movement vibration can introduce errata into the data(person walking nearby or truck driving outside.)

the video you are talking about is a recreation of some Italian location from all the images on flickr. the software doesn't automatically generate the clean logical 3d meshes everyone thinks of, i don't think.

It may be similar to a siggraph 2003 presentation I sat in on. Video footage is used to reverse calculate a depth map which is then used to displace a massive nurbs mesh from a camera projection. The students/researchers who came up with it were using it as a way to archive archaeological excavation data. The mesh was generated and manually tweaked for errors, then the color from source images was projected onto the mesh.