I want to extract 3D information from either two or more images or
video. It is sufficient to have approximated 3D information thus I do
not need a full 3D model, but the depth information is sufficient
(e.g. like disparity maps).
I have implemented/tested several algorithms from epipolar geometry
for uncalibrated stereo pictures but even with hand-marked points from
both images the quality is horrible (at least for non-example images).
So, what is the current state of the art work (papers, demos, etc.)
for extraction of depth information from a) multiple images and b)
Are there any demos/source code out there to try?
I am looking forward to your suggestions, as I do not know how to
continue at the moment.
This is a topic I keep meaning to experiment with... You might try
SIFT or SURF for automated feature extraction and matching. Having
more keypoints may improve the model quality. Also check whether SLAM
techniques apply to your problem, especially for the video processing;
openslam.org may be a good start.
Have you tried the EGT?
Demo yes. Try this one. In swedish but still very butiful.
You have to check the checkbox which reads "I approve of the conditions".
According to my sources this was created by merging data from several
images captured from a low-flying airplane. No laser scanners no
nothing. Just cameras and GPS for tracking plane movements. Result,
complete elevation data and texture for most things. Note that bridges
may look a bit strange as the algorithm does not seem to handle holes
I don't know about Demo's or software but I have used an excellent piece
of video survey software -
Working between images you can measure any point in 3D space to any
other point. Most points (depending on the proximity to camera and
accuracy of click) can be measured with an accuracy of +/-5mm to +/-50mm
we've only requires the nearest 500mm.
I don't think I've seen anything like it, fabulous tool.