tree daac388316aa729d1599439bce6b0a2cad2c8ecb
parent 0292e2e9a5679d815c5d855d6f166b5d4b8a6d20
author yjshen <yjshen@google.com> 1586476353 -0700
committer Damon Shen <yjshen@google.com> 1586893258 +0000

Consider quantization factor in plane-wise filter.

In plane-wise temporal filtering strategy, filter weight is assigned
independently from the quantization factor used in video compression.
However, when we use small q (i.e., with high bitrate), a weaker
filtering strength is expected since we would like to keep more details.

This CL improves the plane-wise strategy by considering the quantization
factor for filter weight assignment. In particular, when q is large
enough (>=16), nothing is changed. When q is less than 16, we will
reduce the filtering strength w.r.t. the q-value. This change
significantly improves the performance on high-bitrate encoding.

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      -0.116     -0.099   -0.040
midres2     -0.047     -0.044   -0.017
hdres       -0.142     -0.197   -0.075
hdres2      -0.010     -0.012   -0.004

Under Speed-1 (two-pass mode):
          avg PSNR   ovr PSNR     SSIM
midres      -0.132     -0.124   -0.040
midres2     -0.057     -0.053   -0.020
hdres       -0.158     -0.177   -0.066
hdres2      -0.014     -0.018   -0.005

STATS_CHANGED

Change-Id: I486e66770a4454fb1f72c3276cd20a4b3ae3dd3d