I've written a piece of software using FindStereoCorrespondenceGC, and while it's generating great disparity maps (in terms of object boundaries), the actual depth quality is mediocre. It's basically put every object onto a 2.5D plane.
The images have all been pre-rectified, although not using openCV. The camera is a Pt. Grey Bumblebee, so I used the Triclops rectification functions.
I was wondering if there's any way to improve the disparity map quality with graph cuts, or do I need to start using a hybrid method.