tree a080e1669248afdcf4d2e0514716b70d063b5838
parent e82c671721bfadf626613ed6341261419ac3d4ac
author yjshen <yjshen@google.com> 1586461362 -0700
committer Damon Shen <yjshen@google.com> 1586837884 +0000

Consider motion search error in plane-wise filter.

In plane-wise temporal filtering strategy, filter weight is assigned
based on motion search. Currently, only a small neighborhood (i.e., a
5x5 window) is considered for each individual pixel. However, the motion
search result of the entire block also reflects the search accuracy.

This CL improves the plane-wise strategy by considering both local (5x5
window) and global (the entire filtering block) information to assign
filter weight. In particular, the window-wise error plays a more
important role than the block-wise error, but both of them are involved
in the filtering process.

NOTE: This CL only affects the performance on midres and hdres datasets.

Experimental results:

Under Speed-4 (two-pass mode):
          avg PSNR   ovr PSNR     SSIM
midres      -0.056     -0.015   -0.026
midres2     -0.029      0.000   -0.000
hdres       -0.085     -0.076   -0.049
hdres2      -0.050     -0.028   -0.017

Under Speed-1 (two-pass mode):
	  avg PSNR   ovr PSNR     SSIM
midres      -0.129     -0.033   -0.036
midres2     -0.074     -0.044   -0.026
hdres       -0.106     -0.091   -0.098
hdres2      -0.070     -0.055   -0.057

STATS_CHANGED

Change-Id: Ie883a0105b1bdf4cd41b40b2e5be7f2f7178dd50