Replace RANSAC by plane fitting

[p0nce]

Is it possible to have most of the visual quality with less CPU?
RANSAC is pretty costly if the plugin depth isn't smooth (early exit).
And 16-bit depth is going to make non-smooth depth map pretty common.
Could a plane fitting method work?

[p0nce]

Curent algo:

• compute distance of every points to 12 potential planes
  • 4 with center and two middle points
  • 8 with center, one middle point, and one diagonal point
• each plane that has more than 3 points on it contribute to the output normal. The more inlier points, the more score, with hand-tuned weights.
  This seems learnable? Paramters would be depth difference to center.
  Also there is two axis of symmetry.

[p0nce]

Some directions badly handled:

or perhaps it's just #430

[p0nce]

Tried least-square plane-fitting:

• Baseline
  
• Plane-fitting
  

[p0nce]

Least-square plane-fitting makes smoother surfaces, which is neat for rotating knobs, but the overall effect is that there is less sharpness. It would be much less expensive in CPU though.

[p0nce]

Adapting the sigma of the gaussian kernel for covariance can increase sharpness, while still being anisotropic => https://observablehq.com/@jobleonard/gaussian-kernel-calculater

[p0nce]

We can totally have something that is both faster and nicer!

[p0nce]

New plane fitting method is faster and better lookign than RANSAC.
This means there is no need for a Z early exit if depth is flat.
Also the normal extraction is isotropic it manages better rotating knobs and circles.
Enabled by default in v12.1.6
RANSAC is removed.

[p0nce]

Maybe a tiny bit more sharpness would be good. Reopened.

[p0nce]

Made a new issue instead.