tree 7bc0c72573e302b33d8f50f0aa028d1ac7d9b452
parent 4895d98bef216814462126209dd08036e7eca6d9
author yjshen <yjshen@google.com> 1578098134 -0800
committer Cheng Chen <chengchen@google.com> 1578362564 +0000

Add dropout optimization.

Dropout optimization can be viewed as a fast mode of trellis
optimization, yet not as accurate. The rational is to forcibly set some
quantized coefficients, which are in the middle of zeros, to zeros. The
main goal is to reduct the bits used for encoding but barely affect the
decoded image. Compared to trellis optimization, this is much faster.

There are two usage of dropout optimization:
(1) A further optimization after trellis optimization.
(2) An alternative to trellis optimization, especially in real-time
mode when trellis optimization is disabled due to high time complexity.

NOTE: For now, dropout optimization is performed to all frames in
real-time mode, and only performed (after trellis optimization) to intra
frames in two-pass mode by default. This is harded coded, and should be
made configurable in the future.

Experimental results:

Under Speed-4 (two-pass mode):
          avg PSNR   overall PSNR     SSIM
lowres       0.008         -0.009    0.029
midres      -0.005         -0.010   -0.000
ugc360p     -0.006         -0.010    0.005

Under Speed-6 (real-time mode):
          avg PSNR   overall PSNR     SSIM
lowres      -0.033         -0.029    0.058
midres      -0.072         -0.135   -0.037
ugc360p     -0.163         -0.218   -0.137

STATS_CHANGED

Change-Id: Icabd73546fefe0d0fd9dc2d112754e84d8c5d550