Improved bilateral filter performance
More than doubled the performance of the filter and correct a bug in the OpenCV version that resulted in flashing from a race condition. Currently the window size is hard coded as 9x9, some experimentation showed this to be a good size with still great performance. Can now run this at 40 iterations.
Further experiments are to include:
- using max colour distance encountered in a direction, I know this reduces bad values (but also density)
- perhaps weight by colour distance with corresponding pixel in right image? If a disparity would result in a high colour inconsistency then probably that should get a low weight?