tree 2ad47105eaea23f0ba2e521dbc43134fdd6efcd5
parent 90222123fa57b18b667b07e669daabae433aab08
author Rachel Barker <rachelbarker@google.com> 1705683702 +0000
committer Rachel Barker <rachelbarker@google.com> 1706221622 +0000

Improve RANSAC algorithm

Implement a few ideas from the literature about RANSAC. Many ideas
were tested, but the ones which proved to be useful are:

* Remove degeneracy checks. The model fitting functions implicitly
  check for situations where the model is underdetermined, and
  inlier counts provide us with a better way to exclude bad models

* Fit initial models using the minimum number of points possible.
  Counterintuitively, this is better than using more points. The reason
  is that we generate a good model if and only if *all* of the points
  in the initial fit are inliers. Selecting fewer points at first raises
  the probability of this happening, and so is more likely to generate
  good models.

* Refine the best detected model multiple times. This helps us find
  a large inlier set, and therefore leads to a better final model.

 Speed | BDRATE-PSNR | BDRATE-SSIM |   Enc time
-------+-------------+-------------+-------------
   1   |   -0.029%   |   -0.030%   |   +0.053%
   2   |   -0.044%   |   -0.042%   |   +0.021%
   3   |   -0.066%   |   -0.050%   |   +0.046%
   4   |   -0.076%   |   -0.084%   |   +0.160%

STATS_CHANGED

Change-Id: I46f04071f6927f50a53ebc60a23503a6e4cf8f64