Photosynth/Seadragon - AMAZING!!

[formatt]

people that are into this kind of stuff are going to be floored...

http://www.ted.com/index.php/talks/view/id/129

Without question, one of the most amazing things I've ever seen. I can't imagine how much work went into this or how you'd even know where to start.

I had to watch this 7 min video a few times because I couldn't believe what I was seeing/hearing. The first thing that popped into my head was the part in Blade Runner where he's looking around inside photos.

[Rhino95]

Wow.

[formatt]

Thats exactly what I said after I watched it the first time... i had to rewatch it a few times to take it all in. insane.

[kPs]

That is nuts!

[Jake_Blues]

I watched this without audio, so I'm not sure what the discussion was, but my impressions from seeing the application:

The first part didn't impress me much, just zooming in and out on static photos, and the second demo (stitching photos together into a panorama) has been done many times before.

The part where they texture-mapped a 3D object with photos was neat, but my gut instinct is you'd need the 3D model first, and it would help to know where the pictures were taken from. You could, with some really clever pattern matching, do it without knowing the location they were taken from.

I could see this becoming interesting if you were to combine a really good Internet search engine that was optimized specifically to find photos, and a really good pattern matching/stitching algorithm. You could get a complete view of anything famous enough that large numbers of people take photos of it. Like, say, J-Lo's booty.

-E

[formatt]

The first part didn't impress me much, just zooming in and out on static photos, and the second demo (stitching photos together into a panorama) has been done many times before.

The part where they texture-mapped a 3D object with photos was neat, but my gut instinct is you'd need the 3D model first, and it would help to know where the pictures were taken from. You could, with some really clever pattern matching, do it without knowing the location they were taken from.

Jake, you're kidding right? Great... now watch it again with audio and hear why it's so impressive. It's slightly more involved than "zooming" and "stitching". :rolleyes:

There were no 3D objects to be textured mapped... that's the point. The pictures were taken from every known source available on the internet and compiled to form a 3D object. Think of it as "Google Images" that builds objects. Images would have metatags (like keywords on websites). You search for a particular subject and it would pull ALL images to form an object.

His example with the postcard was perfect. Say you have a bunch of different photos: a front elevation of the Whitehouse in one photo from a informative site regarding US history. In the background of a family photo from your recent D.C. trip, you have a slightly skewed angle of the Whitehouse that might show the front elevation plus a small portion of the side. Then you have something like a googlemas image of the Whitehouse from the top. Another image from a DC skyline postcard. Etc... The 3D object might be pulling in 50,000+ different images from all over the internet to form this object. Not only that, say some of the photos have closeups of a particular object on the Whitehouse... you can zoom in on that 3D object almost infinitely to see details. The only thing stopping you is the amount of info gathered. PLUS, image DPI is no long an issue... (that's what he was zooming in on with the static images). The only limiting factor now is the dots per inch on your monitor.

[Jake_Blues]

You're going to have to add a whole bunch of information somehow. Metatagging isn't going to be enough, unless it holds a tremendous amount of info.

For instance, what if I want a 3-D model of the statue of liberty, probably one of the most photographed subjects in the world. Map together these images for me:

http://upload.wikimedia.org/wikipedia/commons/thumb/0/00/Paris.seine.liberty.500pix.jpg/200px-Paris.seine.liberty.500pix.jpg

http://upload.wikimedia.org/wikipedia/commons/thumb/4/4f/Statueofliberty.JPG/400px-Statueofliberty.JPG

http://www.roadsideamerica.com/attract/images/nv/NVLVEnyny_2932.jpg

Not to mention people who have photoshopped images, scanned in paintings, etc. It's an admirable goal, but I can't see it working unless the metatagging contains volumes of info.

-E

[formatt]

He explains it all. Those two images you posted are exactly what it's looking for. The more images it can reference, the better. The first images shows how the flame looks from more of an elevation view where at the second image is a lower perspective. It's use both of them to build the 3D object.

[Jake_Blues]

He explains it all. Those two images you posted are exactly what it's looking for. The more images it can reference, the better. The first images shows how the flame looks from more of an elevation view where at the second image is a lower perspective. It's use both of them to build the 3D object.

There is only one problem... that first image I posted is NOT the statue of liberty.

-E

[Yota]

You're going to have to add a whole bunch of information somehow. Metatagging isn't going to be enough, unless it holds a tremendous amount of info.

For instance, what if I want a 3-D model of the statue of liberty, probably one of the most photographed subjects in the world. Map together these images for me:

http://upload.wikimedia.org/wikipedia/commons/thumb/0/00/Paris.seine.liberty.500pix.jpg/200px-Paris.seine.liberty.500pix.jpg

http://upload.wikimedia.org/wikipedia/commons/thumb/4/4f/Statueofliberty.JPG/400px-Statueofliberty.JPG

http://www.roadsideamerica.com/attract/images/nv/NVLVEnyny_2932.jpg

Not to mention people who have photoshopped images, scanned in paintings, etc. It's an admirable goal, but I can't see it working unless the metatagging contains volumes of info.

-E

You need to listen to it. It does this automatically in some way. Pretty damn slick.

It actually doesn't matter much if your image is of the real statue if it is an appropriate facsimile that meets whatever criteria the algorithm sets. In his demo he actually shows an image of a kid standing in front of a poster of Notre Dame - but it met the criteria.

[Trango]

Um, this is called short-range photogrammetry, and even if you don't know much about the camera models ("model" here is a term that involves full characterization of the imaging sensor, optics, and fixture), it's still possible to use simple parallax-type analysis to create a relative (meaning, not absolute, or without certainty as to the scale of the dimension) obliquely-imaged distance measurements and the creation of a terrain map.

So, the technologies happening here are automatic coregistration, mosaicing, color balancing, stretching/warping, and a hell of alot of server-side processing (I can only hope) to accomplish that photogrametry. I'm familiar with the technology, but the quantity of computation would be fairly staggering unless they found a large amount of elegance in the technology.


If the camera models are known, then even 2 cameras can image something and give you full "DEM" or "DTM" information for imaged features, creating a so-called 2D surface map. With a host of photos, you image many more features, and you can get pretty close to full 3d awareness of an object's surface.

And, well, he did explain this in layman's terms... I post the background science behind it. This is, essentially, n-image oblique relative photogrammetry, if I were to have to give it a reference.

I don't know what the complaint is here, but hopefully I've given enough background on this stuff for some usefulness.

[Ric]

thats sweet.. very cool

[Backwoods Rambler]

That was fascinating! :eek::hail:



That is all.........

[formatt]

You need to listen to it. It does this automatically in some way. Pretty damn slick.

It actually doesn't matter much if your image is of the real statue if it is an appropriate facsimile that meets whatever criteria the algorithm sets. In his demo he actually shows an image of a kid standing in front of a poster of Notre Dame - but it met the criteria.

please, yota. you can lead the horse...

Bob.. you're right on track. Although, for that DTM, I'm not entirely sure how you could pull it off with 2 camera points. I'd think that regardless of how the 2 camera points would be oriented, it'd still only give you a very skewed X/Y plane.

I've been going over this stuff in my head all morning... 15 years ago, who would have thought that the internet would be as vast as it is now.. as powerful as it is. IMO, I really don't think that the server end for something like this would be that tough to imagine in the very near future... another 10 years?

[Trango]

2 point photogrammetry (as from stereo pairs) works for this, easily. You have to realize that if you would expect an image to be skewed after capture (and if you know that this is, indeed, the case), you can easily apply a transfer function to take out the skew.

One close range photogrammetry company:
http://www.adamtech.com.au/

[Yota]

Um, this is called short-range photogrammetry, and even if you don't know much about the camera models ("model" here is a term that involves full characterization of the imaging sensor, optics, and fixture), it's still possible to use simple parallax-type analysis to create a relative (meaning, not absolute, or without certainty as to the scale of the dimension) obliquely-imaged distance measurements and the creation of a terrain map.

So, the technologies happening here are automatic coregistration, mosaicing, color balancing, stretching/warping, and a hell of alot of server-side processing (I can only hope) to accomplish that photogrametry. I'm familiar with the technology, but the quantity of computation would be fairly staggering unless they found a large amount of elegance in the technology.


If the camera models are known, then even 2 cameras can image something and give you full "DEM" or "DTM" information for imaged features, creating a so-called 2D surface map. With a host of photos, you image many more features, and you can get pretty close to full 3d awareness of an object's surface.

And, well, he did explain this in layman's terms... I post the background science behind it. This is, essentially, n-image oblique relative photogrammetry, if I were to have to give it a reference.

I don't know what the complaint is here, but hopefully I've given enough background on this stuff for some usefulness.

But do you agree that it's a pretty titsy implementation of all that? :D

[Yota]

Can you picture what would happen if this was done with porn!

What I was thinking the whooole time.

I was also thinking about what would happen if you had some chick posing nekked on one of those mayan temples or somesuch and their software photospatulated it into the mix! :eek:

Or I search for "Toyota SUV" and it photospatulates in some pic that it grabs from some fratboy site showing a dude taking a dump on his buddy's 4Runner?! :lmao:

(photospatulated: trying in vane to sound all smart like Trango :D Were you convinced?)

[jdogg4]

Cool As Hell!

[Guammy]

That's amazing!!!!! :eek:

I loaded the software and have been checking out the photo's they have loaded in their 5 collections on on the website.

http://labs.live.com/Photosynth/default.html

[rch]

Wow, that program is pretty sweet. That would make a cool album of a trails and such.

[Yota]

I played with it. It was kinda fun.

IBGOOGLEBUYSIT