tree 7eac39571986d184aee7a9ac4ed6bdd5d64c5ded
parent ac4ff32713a4f6961f8ef7251eb98497b5512188
author Rachel Barker <rachelbarker@google.com> 1682651288 +0000
committer Rachel Barker <rachelbarker@google.com> 1706710112 +0000

Corner match: Use bidirectional matching

Remember the best match for each feature in both the source and ref
frames. Only generate correspondences for points which are each other's
best matches. This removes a lot of spurious correspondences, which
helps improve the average quality of the generated correspondences.

In addition, rearrange the code to be much more efficient. If there
are n features in each of the source and ref frames, then the matching
loop is run O(n^2) times. Therefore it is beneficial to push as much
work as possible out into pre- and post-processing loops which run
only O(n) times.

Finally, expand patch size from 13x13 to 16x16, as it gains a little
extra quality at minimal cost.

Results @ "good" mode speed 4:

     Compared to      | BDRATE-PSNR | BDRATE-SSIM | Encode time
----------------------+-------------+-------------+-------------
Previous corner match |   -0.044%   |   -0.052%   |   -3.487%
       Disflow        |   +0.003%   |   -0.007%   |   +7.482%

No change to encoder output, as we currently use disflow.

Change-Id: I0abd73717925623eeb3948889eae3ca38c61811d