pouët.net

I just wondering if anyone have heard of such thing where the direction of video camera is extracted from frames of the video itself?

is it possible?

Dance as you learn:

http://www.youtube.com/watch?v=DgGV3l82NTk

No coder left behind!

I suppose you mean match-moving?

Extracted from the metadata or analytically from the video itself? A lot of smartphones include compass/accelerometer/gyro these days, embedding camera location + direction wouldn't be difficult at all.

Not sure what you mean here, but recently real-time marker-less image tracking is getting very reliable, allowing a lot of things to be done, like augmented reality of course, but also fun things like acquisition of terrain in 3D by filming from a plane (I heard Apple bought such a technology, so one can expect them to make some announcement in a not so far future), etc. This is not my field though, so I don't know what are the most relevant links.

Anyway, as long as you have enough info to reconstruct the 3D scene, knowing the direction of the camera at a given frame is just a detail.

Gargaj: well yeah something like match-moving but with out locking on certain object i mean i really don't want some sort pre-setup phase for the system.

psonice: nice idea out there i didn't think about the compass! knowing that most of of recent phones have it.

Zavie: well i am planning on something like augment reality in 3DS and other devices. what i want is to implement a 3D scene withing the real video clip and i don't want to relay much on defining specific object.

maybe i need to dig in real-time marker-less image tracking? if you got any information to share with i would happy :).

Dunno if you noticed but he already said he wants to do it realtime. But A+ for Google skills.

Didn't notice, and I didn't google it.

wasn't this already done in some remote control app for one of those quadcopter thingies?

tried it 1 or 2 years ago. aferteffects has something like that ,and there's a program called voodoo which makes 3d point clouds from the video, I think it can recognize camera motion too. That voodoo stuff works fine with sample frames, but I failed to get good results with a cheap digital camera.

xernobyl: thank you for the link yeah something like what this application do but as Gargaj mentioned i am aiming for realtime.

hooverphonique: have you seen any paper or article that talks about the tech. in it. is so could you share.

pista: sweets i have checked Voodoo site from their description looks like they are using mark-less tracking well i'm downloading it atm hoping that it has something useful in it :).

thanks for the feed,

I dunno if this helps, but you could try to create a 3d model from the pictures you got from the camera and with this 3d information you could try to match features on the current picture. with this approach the user first would need to capture some environment pictures to create the model and afterwards could place 3d objects into this environment.
for some info on creating 3d models from movies google for "structure from motion", dunno what is needed to make this realtime.

there is an argumented reality sdk from qualcomm available for android, but that needs an printed picture as reference and can only handle objects on this flat plane referenced by the printed image. the nintendo 3ds has an augumented reality demo, too, but it also needs a card as reference (to extract the current camera orientation and position).

I guess from your description you're aiming for something a little like this: http://www.augmentedplanet.com/2011/08/ball-invasion-ar-gaming-using-slam-technololgy/

The tech behind it should be easy enough to find out about, and there's enough keywords in that article to get google pointing in the right direction :)

psonice:yeah the idea is similar to the game that mentioned on your link. I feel i am getting close to it :).

well, thank you everyone for your feed back. I'll keep you update about it!

@panic:

Here are keywords for you: PTAM (rather for embedded) and DTAM (rather for high end). Have fun on google. ;)


I meant that those techniques allow to reconstruct some spatial information, so if we can do that, inferring the camera orientation certainly is doable.