Re: Stereo pair reconstruct 3D scene?

Matthias Melcher wrote: 
Nicholas Sherlock wrote:

I want to synthesize a viewpoint somewhere between the two cameras using this information. What's the best way to do this? I could see plotting all of those pixels as cubes or spheres in 3D and pointing a virtual camera at them, but is this really the best way?


You get the view point and viewing angle if you know the distance between the cameras and their opening angle. Be aware though that camera lenses are by no means mathematical projections from a point. You need to measure lens distortion and compensate for it which involves filtering and reduces the quality of you depth information. Also, you can only get depth information in vertical patterns (assuming your cameras are positioned next to each other). So you cube model will have huge gaps on evenly lit surfaces. Lastly, you will very likely find edges that 'look alike' but are actually not the same edge, generating entirely false depth values.

Have fun and let us know how it works out ;-)

Thanks. I'm ignoring lens distortion for the moment (Just to get it working would be nice :)). Depth information can be sort of interpolated between the edges on smooth surfaces. Horizontal patterns can be recognized by using a camera triplet where two cameras are on a horizonal axis and one is a vertical distance above the center of the two lower cameras.

What I really want to know is if the "Cube model" is the best way of displaying the end result?

Check out what these people have done (Look at the movie at the bottom of the page, awesome!):
http://axon.physik.uni-bremen.de/research/stereo/Cyclops/index.html

I guess I could ask them :).

Cheers,
Nicholas Sherlock
.