real-time adaptive image compression: supplementary...

Real-Time Adaptive Image Compression: Supplementary Material

WaveOne JPEG JPEG 2000 WebP BPG

0.96 0.97 0.98 0.99MS-SSIM

5

10

20

40

80

160

320

Tim

e (

ms)

(a) Encode times.

0.96 0.97 0.98 0.99MS-SSIM

5

10

20

40

80

120

Tim

e (

ms)

(b) Decode times.

Figure 1. Average times to encode and decode images from the RAISE-1k 512× 768 dataset. Note our codec was run on GPU.

0.5 1.0 1.5 2.0 2.5 3.0Bits per pixel

0.960

0.965

0.970

0.975

0.980

0.985

0.990

0.995

1.000

MS-S

SIM

Quality 30Quality 40Quality 50Quality 60Quality 70Quality 80Quality 90Uncompressed

Figure 2. We used JPEG to compress the Kodak dataset at various quality levels. For each, we then use JPEG to recompress the images,and plot the resultant rate-distortion curve. It is evident that the more an image has been previously compressed with JPEG, the betterJPEG is able to then recompress it.


ReLU

Bat

chN

orm

Mean

Di�erence

C 3

2, S

2, 3

x3

Di�erence Di�erence Di�erence

Mean Mean Mean

ReLU

Bat

chN

orm

C 6

4, S

1, 3

x3

ReLU

Bat

chN

orm

C 6

4, S

2, 3

x3

ReLU

Bat

chN

orm

C 6

4, S

1, 3

x3

ReLU

Bat

chN

orm

C 6

4, S

2, 3

x3

ReLU

Bat

chN

orm

C 6

4, S

1, 3

x3

ReLU

Bat

chN

orm

C 6

4, S

2, 3

x3

ReLU

C 3

2, S

1, 3

x3

C 32, S 1, 3x3 C 32, S 1, 3x3 C 32, S 1, 3x3 C 32, S 1, 3x3

Rand

omSw

ap

Target

Scalar decision

Reconstruction

Figure 3. The architecture of the discriminator used in our adversarial training procedure. The first module randomly swaps between thetargets and the reconstructions. The goal of the discriminator is to infer which of the two inputs is then the real target, and which is itsreconstruction. We accumulate scalar outputs along branches constructed along the processing pipeline, branched out at different depths.We average these to attain the final value provided to the objective sigmoid function. This multiscale architecture allows aggregatinginformation across different scales. In Section 4 of the main text we discuss the motivation for these architectural choices in more detail.


JPEG

0.0909 BPP

0.1921 BPP

0.4064 BPP

JPEG 2000

0.0847 BPP

0.1859 BPP

0.4002 BPP

WebP

0.1021 BPP

0.1861 BPP

0.4016 BPP

Ours

0.0840 BPP

0.1851 BPP

0.3963 BPP


JPEG

0.0949 BPP

0.1970 BPP

0.4196 BPP

JPEG 2000

0.0941 BPP

0.1953 BPP

0.4069 BPP

WebP

0.1452 BPP

0.1956 BPP

0.4117 BPP

Ours

0.0928 BPP

0.1939 BPP

0.4035 BPP


JPEG

0.1008 BPP

0.2083 BPP

0.3734 BPP

JPEG 2000

0.0953 BPP

0.1939 BPP

0.3690 BPP

WebP

0.1392 BPP

0.1973 BPP

0.3672 BPP

Ours

0.0949 BPP

0.1921 BPP

0.3643 BPP


JPEG

0.1101 BPP

0.2071 BPP

0.4055 BPP

JPEG 2000

0.0947 BPP

0.2014 BPP

0.4002 BPP

WebP

0.1510 BPP

0.1989 BPP

0.4087 BPP

Ours

0.0941 BPP

0.1940 BPP

0.3971 BPP


JPEG

0.1123 BPP

0.2210 BPP

0.4671 BPP

JPEG 2000

0.0994 BPP

0.2183 BPP

0.4638 BPP

WebP

0.1263 BPP

0.2198 BPP

0.4674 BPP

Ours

0.0989 BPP

0.2125 BPP

0.4581 BPP


JPEG

0.0881 BPP

0.1923 BPP

0.4012 BPP

JPEG 2000

0.0846 BPP

0.1889 BPP

0.4002 BPP

WebP

0.0841 BPP

0.1952 BPP

0.4047 BPP

Ours

0.0828 BPP

0.1885 BPP

0.3996 BPP

real-time adaptive image compression: supplementary...

Documents