tree b10a61709c1e2473e5aeb1772575f9a8a88a2e54
parent ad26836eb37a8faf482feb9fd443de6e0dd775f6
author Angie Chiang <angiebird@google.com> 1640741679 -0800
committer Angie Chiang <angiebird@google.com> 1643998375 +0000

Integrate BITRATE_ACCURACY with three pass mode

Preliminary Results:
(There might be some hidden issues that cause  mixed results
in 150 frames)
BITRATE_ACCURACY vs BITRATE_ACCURACY_BL-->av1 vbr mode with 3 pass on

vbr mode 150 frames
metric  avg_psnr ovr_psnr ssim
ugc360p -0.994%  -4.225% -0.606%
lowres   2.395%   1.492%  9.491%
midres   1.499%  -0.089%  3.784%
hdres    0.876%  -0.368%  2.554%

vbr mode 33 frames
metric  avg_psnr ovr_psnr ssim
ugc360p -2.722%  -3.767%  -1.627%
lowres   0.149%  -0.785%   6.130%
midres  -0.806%  -1.378%   1.338%
hdres   -0.869%  -1.620%  -0.198%

vbr mode 17 frames
metric  avg_psnr ovr_psnr ssim
ugc360p -3.684%  -3.529%  -2.729%
lowres   0.491%   0.680%   4.379%
midres  -2.052%  -3.011%  -0.792%
hdres   -1.341%  -2.470%  -0.917%

Notes:
0) The big performance drops in 33 frames and 150 frames are resolved
when using whole sequence's tpl stats.
1) The experiment shows significant gains in ugc360p 150 frames. This is
a good indicator that using shared base layer q index will lead to
better compression performance.
2) We see lowres 150 frames has 1.3% drop, this might be caused by some
hidden issues that required further investigation
3) We didn't report abs_rc_error because my testing process changed.
Haven't figure out how to get abs_rc_error yet. But I looked at several
several examples and figured that we might need to revisit the frame
bits model's accuracy again.

BUG=aomedia:3045

Change-Id: I9c2ac3c4945735c4d2ae8a9a08b91966c0978992