You can search around for algorithms on 'fast image tracking'. You can also look for a library that abstracts over OpenCV's pre-existing object detection or, depending on your needs, implement your own using OpenCV's example: see the 'find_obj.cpp' code in the samples directory of the latest distribution for a guide.

I plan to release the source for this project as soon as the code is good and ready. Keep an eye on theartvertiser.com and/or julianoliver.com/blog for details about this!

Thanks for the links.

It's not so similar really. Bryan Chung is using the widely popular ARToolKit tracker with 'fiducial markers' - in other words specially defined black and white images with a large black border - on which to composite video. The image in the first video he is using is the default marker shipped with the software, 'Hiro'.

The video example I posted here uses any complex image: street advertising, a picture t-shirt or billboards being good examples of candidates. The technique is known as Natural Feature Tracking and doesn't require using pre-defined codes/markers for augmentation to be possible..

Cheers!

Thank you for the explaniton.I learned many technical terms from your reply.Yes your code and software is more complex and more successful but idea is the same,right?But anyway its really good.Selecting TED intro as sample video is another good point :)
Cheers
Erdem

The tracker part started out with OpenCV's own object detection tools (above comment). Later I found the 2d->3d homographic stuff in (OpenCV based) Bazar to be really clean and useful so I decided to go with this over my own (hairier) solutions.

Overall though I've found Bazar to be too slow for AR, especially on a netbook target. The geometric calibration part needs a lot of careful attention to generate a good augmentation also. Since I've found several areas in which to optomise to bring it all up to ~25fps: beforehand it was ~10-15fps on a pretty powerful box..

I'd reccommend just rolling your own solution around the OpenCV planar object detection code if your needs are simple. If you need all that homographic goodness then Bazar is the way to go!

Cheers (and I like the card tricks ;)