search

huzi96 / Coarse2Fine-PyTorch

70 stars 6 forks source link

Require for the RD points used in TABLE 4 #2

Closed ZhangYuef closed 3 years ago

ZhangYuef commented 3 years ago

Thanks for sharing the work!

I am trying to reproduce the results in TABLE 4 of TPAMI 2021 paper. Specifically, I use the open-sourced RD points from tensorflow/compression but I get the different BD-Rate results from TABLE 4. Parts of my reproduce: Methods Kodak Kodak-TPAMI 2021
ICLR2018_factorized 32.76% 32.45%
ICLR2018_hyperprior 6.53% 3.43%
ICLR19 -3.43% -4.94%
BPG 0 0
JPEG 111.24% 115.05%

The second column is my reproduce and the third one is from the paper. I follow the instruction in the paper to use bpp range [0.4, 1.15] for Kodak.

Does the jitter < 5% in RD-Rate can be neglected since those RD curves are got by mathematically fitting? Also, I think it would be greatly appreciated that you could open-sourcing the RD points you used for those methods.

huzi96 commented 3 years ago

The jitter may be due to the difference in the used interpolation method.

Here are the points: [bpp, PSNR (dB)] ICLR2018 factorized: [[ 0.12004768, 26.81590579], [ 0.19721815, 28.30260532], [ 0.31842126, 29.99836244], [ 0.49367947, 31.72285964], [ 0.73317295, 33.6688382 ], [ 1.04842122, 35.77904641], [ 1.46034749, 37.99930683], [ 1.96733093, 40.0629516 ]]

ICLR2018_hyperprior: [[ 0.12748549, 27.24206466], [ 0.20631663, 29.06472367], [ 0.3211814 , 30.89077978], [ 0.48136393, 32.73602189], [ 0.68496789, 34.63703971], [ 0.96489122, 36.86098719], [ 1.31052992, 38.83483582], [ 1.73873901, 40.82808222]]

ICLR19: [[ 0.1005444 , 27.64089224], [ 0.14531793, 28.8015183 ], [ 0.20252567, 29.89431261], [ 0.29850176, 31.19881032], [ 0.43401506, 32.85073809], [ 0.69004907, 35.17374293], [ 0.99436951, 37.31821059], [ 1.27113512, 38.9376673 ], [ 1.59261661, 40.3777134 ]]

BPG: [[ 1.92971208, 41.47887286], [ 1.5989134 , 40.21253023], [ 1.31941054, 38.90195768], [ 1.0802002 , 37.57484552], [ 0.86662716, 36.22240718], [ 0.68469662, 34.85747379], [ 0.53313107, 33.53685383], [ 0.40522512, 32.22686496], [ 0.30310228, 30.97433846], [ 0.22347599, 29.80391885], [ 0.16102346, 28.68154931]]

JPEG: [[ 0.25520409, 26.67051871], [ 0.44836595, 29.13867092], [ 0.60891469, 30.49278333], [ 0.74152035, 31.41972802], [ 0.86686452, 32.17442926], [ 1.0021091 , 32.90818387], [ 1.20890469, 33.915194 ], [ 1.5424313 , 35.36971615], [ 2.7398368 , 39.16140764]]

ZhangYuef commented 3 years ago

The jitter may be due to the difference in the used interpolation method.

Here are the points: [bpp, PSNR (dB)] ICLR2018 factorized: [[ 0.12004768, 26.81590579], [ 0.19721815, 28.30260532], [ 0.31842126, 29.99836244], [ 0.49367947, 31.72285964], [ 0.73317295, 33.6688382 ], [ 1.04842122, 35.77904641], [ 1.46034749, 37.99930683], [ 1.96733093, 40.0629516 ]]

ICLR2018_hyperprior: [[ 0.12748549, 27.24206466], [ 0.20631663, 29.06472367], [ 0.3211814 , 30.89077978], [ 0.48136393, 32.73602189], [ 0.68496789, 34.63703971], [ 0.96489122, 36.86098719], [ 1.31052992, 38.83483582], [ 1.73873901, 40.82808222]]

ICLR19: [[ 0.1005444 , 27.64089224], [ 0.14531793, 28.8015183 ], [ 0.20252567, 29.89431261], [ 0.29850176, 31.19881032], [ 0.43401506, 32.85073809], [ 0.69004907, 35.17374293], [ 0.99436951, 37.31821059], [ 1.27113512, 38.9376673 ], [ 1.59261661, 40.3777134 ]]

BPG: [[ 1.92971208, 41.47887286], [ 1.5989134 , 40.21253023], [ 1.31941054, 38.90195768], [ 1.0802002 , 37.57484552], [ 0.86662716, 36.22240718], [ 0.68469662, 34.85747379], [ 0.53313107, 33.53685383], [ 0.40522512, 32.22686496], [ 0.30310228, 30.97433846], [ 0.22347599, 29.80391885], [ 0.16102346, 28.68154931]]

JPEG: [[ 0.25520409, 26.67051871], [ 0.44836595, 29.13867092], [ 0.60891469, 30.49278333], [ 0.74152035, 31.41972802], [ 0.86686452, 32.17442926], [ 1.0021091 , 32.90818387], [ 1.20890469, 33.915194 ], [ 1.5424313 , 35.36971615], [ 2.7398368 , 39.16140764]]

Thanks for sharing!