Semi-Global Matching (SGM)

SGM is a stereo disparity algorithm for depth estimation.

It is a better way of doing Stereo Matching.

Introduced at NVIDIA.

Uses pixel-level matching.

Implementations

• https://github.com/fixstars/libSGM/tree/master
• https://docs.nvidia.com/vpi/group__VPI__StereoDisparityEstimator.html by VPI
• https://www.mathworks.com/help/vision/ref/disparitysgm.html by MATLAB

Why do we need to specify MinDisparity and MaxDisparity?

the algorithm limits its search for corresponding pixels to within this specified range. This optimization prevents the algorithm from wasting computational resources on searching across the entire width of the image, thereby improving efficiency and speed.

In MatLab, they say (-N,N). Usually it’s like [0, 128). If it’s negative, it’s like the thing actually moved to the left lol. Which wouldn’t make sense.