Adding a 32-point lapping filter

Multimedia / Daala - Nathan E. Egge [dgql.org] - 11 March 2015 22:41 UTC

Added a 32-point lapping filter optimized for 1D coding gain using AR95. Updated the coefficient ordering for the 32x32 bands by training on data collected from subset3 using the new 32-point lapping. Tuned the quantization matrix for the 10 bands in a 32x32 block.

subset1:

RATE (%) DSNR (dB) PSNR -0.01551 0.00091 PSNRHVS -0.33888 0.02433 SSIM -0.81205 0.03764 FASTSSIM -2.49985 0.08521

subset3:

RATE (%) DSNR (dB) PSNR -0.02078 0.00120 PSNRHVS -0.40810 0.02858 SSIM -0.93040 0.04233 FASTSSIM -2.68366 0.10445

video1-short:

RATE (%) DSNR (dB) PSNR -3.15239 0.14884 PSNRHVS -1.08304 0.07259 SSIM -2.64729 0.10882 FASTSSIM -1.75571 0.09164

ntt-short-1:

RATE (%) DSNR (dB) PSNR -2.24747 0.08457 PSNRHVS -0.07304 0.00454 SSIM -1.79033 0.05243 FASTSSIM -1.42222 0.04681

a19b783 Adding a 32-point lapping filter.
src/filter.c | 556 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
src/filter.h | 5 +-
src/pvq.c | 2 +-
src/zigzag32.c | 384 +++++++++++++++++++-------------------
4 files changed, 746 insertions(+), 201 deletions(-)

Upstream: git.xiph.org

  • Share