Date: June 2, 2021.
Here, we give more explanation about the parameters that we used for our plots in Figure 8. (a) and (b) in the main paper.
In Figure 8. (a) and (b), we used MSE
and MS-SSIM
distortion functions for neural network-based model trainings, respectively.
For Balle-extended models, we scale both distortion functions by a factor of (255**2) in order to avoid choosing very small lambda values to operate in the range of 0-0.3 BPP region.
Note that we used window_size=7
in MS-SSIM
distortion function as our training image size is of dimension 128x256.
We use the following lambda values to obtain both Figure 8. (a) and (b):
ours+Ballé2017
:- For
MSE
trainings: [2e-05, 4e-05, 0.0002, 0.0005, 0.0008, 0.0012, 0.0014, 0.0016, 0.002, 0.0026, 0.0032, 0.0038, 0.0044, 0.005, 0.0056] - For
MS-SSIM
trainings: [1.2e-05, 3e-05, 4.5e-05, 6e-05, 8e-05, 9e-05, 0.0001, 0.00014, 0.0002, 0.0008, 0.0014, 0.005]
- For
ours+Ballé2018
:- For
MSE
trainings: [0.0002, 0.0008, 0.0011, 0.0014, 0.002, 0.0026, 0.0032, 0.0038, 0.0044, 0.0048] - For
MS-SSIM
trainings: [1.2e-05, 1.6e-05, 3e-05, 4.5e-05, 6e-05, 0.0001, 0.00014, 0.00018, 0.0022]
- For
Ballé2018
:- For
MSE
trainings: [0.0002, 0.0005, 0.0009, 0.001, 0.0011, 0.0014, 0.002, 0.0026, 0.0032, 0.0038, 0.0044, 0.0048] - For
MS-SSIM
trainings: [1.6e-05, 8e-05, 4e-06, 8e-06, 1.2e-05, 2e-05, 3e-05, 4.5e-05, 6e-05, 7e-05, 9e-05, 0.0001, 0.00014, 0.00018, 0.00022, 0.00026]
- For
Ballé2017
:- For
MSE
trainings: [0.0002, 0.0004, 0.0008, 0.0014, 0.002, 0.0026, 0.0032, 0.0036, 0.0044, 0.0048] - For
MS-SSIM
trainings: [4e-06, 1.2e-05, 3e-05, 3.7e-05, 4.5e-05, 6e-05, 9e-05, 0.00012, 0.00014, 0.00016, 0.00018]
- For
Note that not all lambda values indicated above are shown in Figure 8. (a) and (b).
To carry out JPEG 2000
and BPG
evalutions, we used CompressAI
library. We varied the quality metric, -q
, in the following ranges for BPG
and JPEG 2000
:
- For
BPG
: [10, 11, 12, ... 51] - For
JPEG 2000
: [20, 25, 30, 35, 40, 50, 75, 100, 125, ... 250]