tree 330db74735982deb9a909c3461b3a2bb2c4bd722
parent 7d3955f6feefc8b3cc5dc4a4766b1d561c195753
author Paul Wilkins <paulwilkins@google.com> 1569492222 +0100
committer Jingning Han <jingning@google.com> 1570204849 +0000

Change KF Boost calculation.

Adjust the KF boost calculation to better match the calculation in VP9.
This monitors the difference between the last frame and second last
frame error and uses this to estimate the rate of decay in prediction
quality and ultimately trigger a break out in the boost calculation.

There are significant net benefits for all sets apart from the midres set
where the results are mixed (better vmaf worse ssim). In this set all
the clips are scaled from larger sources and this may be a factor. This
may point to a fruitful area for further changes.

In general there is a decline in boost for very easy clips and the worst
regression case is midres Snow_Mnt_480P which is almost static. In
contrast boost is increased a little for some harder clips like Husky and
ParkJoy.

Results:- (Overall PSNR, SSIM, PSNR-HVS, VMAF) -ve better.

Low Res -0.923	-0.119	-0.974	-1.878
UGC 360 -0.871	-0.608	-0.746	-1.203
Mid Res  0.080	 0.800	-0.133	-0.678
Hd Res  -0.505	-0.405	-0.573	-0.459

As expected this patch requires tweaks to some unit tests that do single
point sanity psnr checks without accounting for size. Hence any reduction
in KF boost may trigger them.

STATS_CHANGED

Change-Id: I6d1239bb6fb6fe7cd5e40a85787022edb03c738f