Opinion on 3D point cloud from disparity map?

Hello,

I work on a stereo imaging/3D reconstrcution algorithm, I computed the 3D point cloud from the disparity map.

Here is the pair of stereo images before and after rectification.

Here is the disparity map (the max value of disparity is 1360) :

And here is the 3D point cloud :

We can not indentified the original scene with the point cloud so my question is what I should/shoudn't do to display correctly the scene in 3D ? I made a simple scene with few big objects to avoid to much little details and simplify the computing but it doesn't seem to be good (by the way I compute the matched points between stereo images and it display only 4-5 matched points... I think it is very few, so maybe the starting images are not good ?)

I would like to know every opinion/advice on my work. If you need to see some part of my code (very simple, made from Matlab examples) just ask :)

Thank you very much for the help !

Answer 1

[Sai Bhargav Avula]

Hi,

The scene has very little texture. Try adding texture which will aid in better 3D recontruction. Also try to recalibrate the stereo camera using MATLAB Stereo calibration app(link). This app directly calculates the reprojection error and epipolar error through which you can access the calibration accuracy.

The following link shows the detailed workflow for creating 3D point cloud from two stereo images

https://www.mathworks.com/help/vision/examples/depth-estimation-from-stereo-video.html

For rectifying the stereo parameters follow the link below

https://www.mathworks.com/help/vision/ref/rectifystereoimages.html