Back to Search Document details
17th Meeting: Brussels, January 2020 2020-01-10 16:28
CE2: Summary Report on Palette Mode Coding
Abstract
This contribution provides a summary report of Core Experiment 2 on palette mode coding. 5 tests have been tested in CE2 in between JVET-P and JVET-Q meetings, to study and evaluate technologies related to palette mode. In this report, coding performance and complexity of these tests are reported and analyzed. In particular, test results against VTM-7.0 anchors (4:4:4 settings) are provided to show the coding efficiency and complexity trade-off of each tool. Some of the tests are also evaluated under low QPs testing conditions. Crosschecking results for the performed tests are integrated in this contribution.
JVET-Q0022 CE2: Summary Report on Palette Mode Coding [X. Xu, Y.-H. Chao, Y.-C. Sun, J. Xu]

This was initially discussed on 7 January at 1900-2150.

Test

Tester

Document

Tool description

Cross checker

CE2-1.1

W. Zhu

(ByteDance)

JVET-Q0075

QP dependent fixed-length binarization for escape coding

D. Luo

(Alibaba)

CE2-1.2

H.-J. Jhu

(Kwai)

JVET-Q0066

Escape value is binarized by fixed-length code with a length determine by maximum quantized value no change on escape de-quantization

Y.-H. Chao

(Qualcomm)

CE2-1.3

H.-J. Jhu

(Kwai)

S. Yoo

(LGE)

JVET-Q0067

Escape value is binarized by truncated binary code with a length determine by maximum quantized value

Y.-H. Chao

(Qualcomm)

CE2-2.1

W. Zhu

(ByteDance)

H. Jang

(LGE)

JVET-Q0076

Resetting predictor palette at CTU row

M.G. Sarwer

(Alibaba)

CE2-3.1a

Y.-H. Chao

(Qualcomm)

JVET-Q0064

Simplified palette predictor update process for small blocks: restrict entries allowed for prediction to the first W x H entries

H.-J. Jhu

(Kwai)

CE2-3.1b

Y.-H. Chao

JVET-Q0064

Simplified palette predictor update process for small blocks: restrict entries allowed for prediction to the first (W x H ) / 2 entries

H.-J. Jhu

(Kwai)

CE2-3.1c

Y.-H. Chao

JVET-Q0064

Simplified palette predictor update process for small blocks: Bypass predictor update for CU size < 8x8

H.-J. Jhu

(Kwai)

CE2-3.1d

Y.-H. Chao

JVET-Q0064

Simplified palette predictor update process for small blocks: Bypass predictor update for CU size <= 8x8

H.-J. Jhu

(Kwai)

CE2-1.x is on escape coding modification. Results:

CE2-1.x test results with VTM-7.0+DualTree=1, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-1.1

0.07%

0.00%

0.01%

99%

100%

0.07%

-0.06%

-0.06%

99%

100%

CE2-1.2

-0.09%

-0.07%

-0.06%

101%

100%

-0.03%

-0.06%

-0.04%

100%

100%

CE2-1.3

-0.10%

-0.07%

-0.05%

97%

99%

-0.05%

-0.11%

-0.06%

100%

102%

TGM 720p

CE2-1.1

0.00%

-0.01%

0.07%

100%

99%

0.05%

0.00%

-0.09%

100%

100%

CE2-1.2

-0.01%

-0.08%

-0.02%

100%

100%

0.01%

-0.01%

0.09%

100%

100%

CE2-1.3

-0.02%

-0.08%

-0.05%

97%

96%

0.00%

0.00%

0.09%

100%

100%

Animation

CE2-1.1

-0.01%

0.00%

0.02%

99%

99%

0.07%

0.07%

0.06%

99%

99%

CE2-1.2

-0.03%

-0.01%

0.03%

101%

99%

0.05%

0.03%

0.11%

99%

100%

CE2-1.3

-0.03%

-0.03%

0.00%

98%

95%

0.00%

0.00%

0.04%

100%

101%

Mixed Content

CE2-1.1

0.05%

0.04%

-0.01%

99%

99%

0.11%

-0.01%

0.02%

100%

100%

CE2-1.2

-0.04%

0.02%

-0.03%

100%

100%

0.14%

0.29%

0.28%

100%

100%

CE2-1.3

-0.03%

0.01%

-0.02%

98%

97%

0.06%

0.14%

0.20%

100%

99%

Camera-Captured

CE2-1.1

-0.01%

-0.01%

0.00%

100%

99%

0.01%

0.00%

0.04%

99%

99%

CE2-1.2

0.00%

-0.01%

0.00%

100%

101%

0.01%

-0.03%

0.02%

100%

101%

CE2-1.3

0.00%

-0.01%

0.01%

98%

97%

0.01%

-0.03%

0.02%

100%

102%

CE2-1.x test results with VTM-7.0+DualTree=0, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-1.1

-0.02%

-0.18%

-0.16%

100%

100%

-0.10%

-0.38%

-0.29%

100%

99%

CE2-1.2

-0.20%

-0.08%

-0.04%

100%

99%

-0.15%

0.04%

-0.09%

100%

100%

CE2-1.3

-0.24%

-0.22%

-0.16%

98%

98%

-0.16%

0.02%

-0.09%

100%

101%

TGM 720p

CE2-1.1

-0.02%

-0.03%

-0.03%

100%

100%

-0.06%

-0.09%

-0.28%

100%

99%

CE2-1.2

-0.04%

-0.03%

-0.01%

101%

100%

-0.03%

0.05%

-0.02%

100%

100%

CE2-1.3

-0.07%

-0.03%

-0.05%

97%

96%

-0.03%

0.09%

0.03%

100%

102%

Animation

CE2-1.1

0.01%

-0.02%

-0.04%

100%

100%

0.08%

0.08%

0.04%

100%

100%

CE2-1.2

-0.01%

0.04%

0.00%

100%

100%

0.09%

0.05%

0.07%

100%

100%

CE2-1.3

-0.04%

0.01%

-0.02%

97%

98%

0.07%

0.09%

0.09%

100%

102%

Mixed Content

CE2-1.1

-0.08%

-0.16%

-0.21%

100%

99%

-0.20%

-0.42%

-0.57%

99%

99%

CE2-1.2

-0.05%

0.02%

0.03%

100%

100%

0.01%

0.09%

0.01%

99%

100%

CE2-1.3

-0.06%

-0.02%

-0.02%

97%

96%

-0.03%

0.13%

0.01%

100%

101%

Camera-Captured

CE2-1.1

0.00%

0.00%

0.00%

100%

99%

0.00%

0.00%

-0.01%

99%

100%

CE2-1.2

0.00%

0.00%

0.00%

100%

101%

0.00%

0.00%

-0.01%

101%

100%

CE2-1.3

0.00%

0.00%

0.00%

98%

96%

0.00%

0.00%

-0.01%

100%

100%

CE2-1.x test results with VTM-7.0+DualTree=1, low QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-1.1

-1.10%

-1.46%

-1.54%

100%

100%

-1.28%

-1.58%

-1.48%

100%

99%

CE2-1.2

-1.40%

-1.26%

-1.31%

101%

100%

-1.23%

-1.23%

-1.25%

100%

100%

CE2-1.3

-1.47%

-1.32%

-1.38%

97%

99%

-1.42%

-1.51%

-1.48%

100%

101%

TGM 720p

CE2-1.1

-0.57%

-0.63%

-0.50%

100%

100%

-0.73%

-0.29%

-0.39%

100%

100%

CE2-1.2

-0.46%

-0.42%

-0.44%

100%

100%

-0.30%

-0.19%

-0.34%

100%

100%

CE2-1.3

-0.54%

-0.47%

-0.46%

98%

96%

-0.28%

-0.20%

-0.32%

100%

101%

Animation

CE2-1.1

-0.33%

-0.33%

-0.38%

99%

100%

-0.21%

-0.39%

-0.37%

99%

99%

CE2-1.2

-0.31%

-0.18%

-0.17%

100%

100%

-0.24%

-0.34%

-0.33%

100%

100%

CE2-1.3

-0.42%

-0.26%

-0.32%

98%

99%

-0.24%

-0.37%

-0.38%

102%

102%

Mixed Content

CE2-1.1

-1.02%

-1.38%

-1.47%

100%

100%

-0.84%

-0.07%

-0.16%

100%

99%

CE2-1.2

-0.94%

-1.08%

-1.13%

100%

99%

-0.39%

0.30%

0.24%

100%

100%

CE2-1.3

-0.96%

-1.12%

-1.15%

98%

98%

-0.52%

0.29%

0.25%

100%

103%

Camera-Captured

CE2-1.1

0.00%

-0.01%

-0.02%

99%

100%

-0.01%

-0.01%

0.02%

99%

100%

CE2-1.2

0.00%

0.00%

-0.02%

101%

99%

-0.01%

-0.01%

0.03%

100%

101%

CE2-1.3

0.00%

0.00%

-0.02%

97%

99%

-0.01%

-0.01%

0.04%

103%

105%

CE2-1.x test results with VTM-7.0+DualTree=0, low QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-1.1

-2.80%

-3.63%

-3.46%

100%

100%

-2.08%

-2.45%

-2.27%

99%

98%

CE2-1.2

-2.58%

-3.03%

-2.93%

99%

101%

-1.80%

-1.95%

-1.84%

100%

100%

CE2-1.3

-2.83%

-3.28%

-3.19%

98%

102%

-2.10%

-2.27%

-2.17%

100%

100%

TGM 720p

CE2-1.1

-1.06%

-1.13%

-1.33%

100%

98%

-0.75%

-0.95%

-0.99%

99%

98%

CE2-1.2

-0.91%

-0.79%

-1.13%

101%

99%

-0.42%

-0.32%

-0.57%

100%

100%

CE2-1.3

-0.93%

-0.84%

-1.16%

98%

97%

-0.52%

-0.43%

-0.69%

100%

102%

Animation

CE2-1.1

-0.36%

-1.27%

-1.22%

99%

100%

-0.31%

-0.51%

-0.56%

100%

98%

CE2-1.2

-0.37%

-0.63%

-0.69%

100%

101%

-0.29%

-0.36%

-0.40%

100%

100%

CE2-1.3

-0.52%

-0.87%

-0.92%

98%

100%

-0.34%

-0.47%

-0.47%

102%

105%

Mixed Content

CE2-1.1

-1.75%

-2.35%

-2.38%

99%

98%

-1.09%

-1.43%

-1.33%

99%

100%

CE2-1.2

-1.39%

-1.74%

-1.71%

100%

100%

-0.63%

-0.55%

-0.49%

99%

101%

CE2-1.3

-1.48%

-1.81%

-1.78%

99%

96%

-0.79%

-0.65%

-0.59%

101%

102%

Camera-Captured

CE2-1.1

0.00%

0.00%

-0.01%

100%

100%

0.00%

0.00%

0.02%

99%

100%

CE2-1.2

0.00%

0.01%

0.00%

101%

99%

-0.01%

-0.01%

0.00%

100%

101%

CE2-1.3

0.00%

0.01%

0.00%

98%

98%

0.00%

-0.01%

0.01%

104%

105%

The benefit at normal QP range is negligible.

For low QP, the BD rates indicate higher gain. From the RD plots, this is mostly contributed at the very low end, where palette escape coding would be lossless for QP=2. However, BD rates may be misleading here, as the PSNR values are extremely high with large fluctuations over the sequence, such that the averaging of PSNR in the BD computation is misleading (and still, gains in the range of 2% are relatively low for screen content classes).

The current results therefore do not allow a real quantification of the benefit.

It was requested to report results that compute BD rate for low QP range based on sequence PSNR (i.e. compute average MSE and then PSNR), as e.g. suggested in JVET-P0393. Such results were made available in v5 of JVET-Q0022. It turned out that the results are similar.

The current method of EG3 coding of escape is likely not optimum for low QP. On the other hand, the range of binarization is made QP dependent in the CE2-1.x proposals, which is also undesirable (as currently QP is currently not required in parsing, such that in particular with local QP adaptation this is highly undesirable and should not be adopted by design principle. It is mentioned that there are non-CE contributions that resolve this issue.

Beyond that, it is also suggested to study the potential benefit of palette for lossless coding of screen content. It could be the case that the impact of a modified escape coding would be even more apparent in the lossless mode.

CE2-2.1 (JVET-Q0076)

Predictor palette is reset at the beginning of decoding each CTU row to enable parallel processing. The setting of using multiple tile columns are used to test to evaluate the performance changes caused by resetting. Furthermore, a picture is divided into 4 tile columns. The version “*” has 4 tile columns, the other version 1 tile column.

CE2-2.1 test results with VTM-7.0+DualTree=1, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-2.1

0.09%

0.11%

0.08%

99%

99%

0.36%

0.52%

0.46%

99%

99%

CE2-2.1*

0.37%

0.50%

0.34%

99%

100%

0.47%

0.64%

0.52%

99%

100%

TGM 720p

CE2-2.1

0.35%

0.06%

0.11%

100%

99%

0.37%

0.42%

0.64%

99%

98%

CE2-2.1*

0.33%

0.20%

0.43%

99%

101%

0.41%

0.42%

0.67%

100%

101%

Animation

CE2-2.1

0.10%

0.25%

0.32%

98%

98%

-0.03%

0.22%

0.36%

98%

99%

CE2-2.1*

0.08%

0.29%

0.23%

100%

100%

0.00%

0.18%

0.33%

100%

100%

Mixed Content

CE2-2.1

0.07%

0.41%

0.31%

98%

99%

0.21%

0.57%

0.53%

98%

100%

CE2-2.1*

0.16%

0.46%

0.38%

99%

101%

0.27%

0.71%

0.64%

100%

101%

Camera-Captured

CE2-2.1

0.00%

-0.01%

0.00%

98%

100%

-0.02%

-0.04%

0.03%

98%

100%

CE2-2.1*

0.01%

-0.03%

0.00%

97%

98%

0.00%

0.01%

-0.01%

101%

101%

"*" refer to using 4 tile columns per picture setting.

CE2-2.1 test results with VTM-7.0+DualTree=0, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-2.1

0.74%

1.09%

0.98%

100%

99%

0.53%

0.85%

0.69%

99%

99%

CE2-2.1*

1.90%

2.34%

2.15%

100%

101%

0.94%

1.20%

1.03%

100%

100%

TGM 720p

CE2-2.1

1.00%

1.43%

2.05%

97%

95%

0.76%

1.25%

1.96%

98%

96%

CE2-2.1*

1.36%

1.97%

2.08%

97%

98%

1.14%

1.71%

2.30%

100%

99%

Animation

CE2-2.1

0.77%

1.71%

1.75%

98%

100%

0.14%

0.74%

0.93%

98%

98%

CE2-2.1*

0.71%

1.47%

1.60%

99%

100%

0.10%

0.62%

0.76%

101%

101%

Mixed Content

CE2-2.1

0.67%

0.90%

0.68%

98%

98%

0.52%

0.92%

0.70%

97%

100%

CE2-2.1*

1.09%

1.40%

1.35%

100%

101%

0.72%

1.03%

0.73%

100%

101%

Camera-Captured

CE2-2.1

0.00%

0.00%

0.00%

99%

102%

0.00%

0.00%

0.00%

97%

98%

CE2-2.1*

0.00%

0.00%

0.00%

100%

101%

0.00%

0.00%

0.00%

101%

101%

"*" refer to using 4 tile columns per picture setting.

The reset is done as for WPP, but the proposal is doing it always (even if WPP is not used). Benefit is not clear, it causes some loss under CTC, and it could also be done as encoder choice (see JVET-Q0502).

CE2-3.x (JVET-Q0064)

Two methods are tested to simplify the palette predictor update for small coding units in order to solve the issue of pipeline latency. In the first method, the elements in the palette predictor that are allowed to be used in predicting palette table is restricted to the first W x H elements (W and H stands for width and height of the CU). In the second method, palette predictor update is bypasses for CU size < T.

Subtests:

  • Restrict the entries in the palette predictor that are allowed to be used in prediction to the first W x H elements
  • Restrict the entries in the palette predictor that are allowed to be used in prediction to the first (W x H) / 2 elements
  • Palette predictor update is bypassed for CU size (W x H ) < 64
  • Palette predictor update is bypassed for CU size (W x H ) <= 64

CE2-3.1x test results with VTM-7.0+DualTree=1, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-3.1a

-0.04%

-0.01%

-0.01%

100%

101%

0.01%

0.00%

0.02%

100%

100%

CE2-3.1b

-0.06%

-0.01%

-0.02%

100%

102%

0.07%

0.06%

0.10%

101%

101%

CE2-3.1c

0.05%

0.04%

0.04%

100%

105%

0.12%

0.14%

0.13%

100%

101%

CE2-3.1d

0.15%

0.09%

0.09%

100%

102%

0.24%

0.30%

0.24%

101%

101%

TGM 720p

CE2-3.1a

0.02%

-0.01%

0.05%

99%

102%

-0.01%

-0.01%

0.05%

100%

98%

CE2-3.1b

0.04%

-0.04%

-0.03%

100%

102%

0.09%

0.07%

0.08%

101%

99%

CE2-3.1c

0.04%

-0.01%

0.07%

100%

101%

0.07%

0.08%

0.09%

100%

102%

CE2-3.1d

0.14%

-0.04%

0.10%

100%

100%

0.19%

0.13%

0.22%

100%

101%

Animation

CE2-3.1a

0.01%

0.01%

0.00%

100%

102%

0.00%

0.03%

0.00%

100%

99%

CE2-3.1b

-0.01%

0.03%

0.02%

101%

102%

-0.05%

0.00%

-0.02%

100%

99%

CE2-3.1c

0.02%

0.06%

0.07%

100%

100%

0.05%

0.07%

0.06%

100%

101%

CE2-3.1d

0.05%

0.11%

0.14%

101%

98%

0.01%

0.02%

0.09%

100%

99%

Mixed Content

CE2-3.1a

-0.02%

0.01%

0.01%

100%

100%

0.07%

0.11%

0.14%

100%

100%

CE2-3.1b

0.01%

0.01%

-0.01%

100%

102%

0.09%

0.10%

0.22%

101%

100%

CE2-3.1c

0.02%

0.05%

0.02%

100%

99%

0.09%

0.12%

0.22%

99%

100%

CE2-3.1d

0.06%

0.08%

0.06%

100%

101%

0.13%

0.13%

0.16%

102%

101%

Camera-Captured

CE2-3.1a

0.00%

0.00%

0.00%

100%

107%

0.00%

0.00%

0.00%

101%

99%

CE2-3.1b

0.00%

0.00%

0.00%

101%

104%

0.00%

0.00%

0.00%

101%

100%

CE2-3.1c

0.00%

-0.02%

-0.01%

100%

104%

-0.01%

-0.01%

0.03%

101%

101%

CE2-3.1d

0.00%

-0.02%

-0.01%

101%

101%

-0.01%

0.00%

0.00%

101%

102%

CE2-3.1x test results with VTM-7.0+DualTree=0, normal QP range

 

 

AI Over VTM-7.0

RA Over VTM-7.0

 

Test#

Y

U

V

EncT

DecT

Y

U

V

EncT

DecT

TGM 1080p

CE2-3.1a

0.08%

0.09%

0.12%

101%

100%

0.06%

0.06%

0.05%

100%

99%

CE2-3.1b

0.47%

0.48%

0.48%

102%

100%

0.33%

0.33%

0.30%

101%

100%

CE2-3.1c

0.13%

0.19%

0.16%

101%

102%

0.12%

0.17%

0.13%

100%

100%

CE2-3.1d

0.41%

0.55%

0.47%

101%

99%

0.38%

0.49%

0.40%

101%

100%

TGM 720p

CE2-3.1a

0.04%

0.00%

0.05%

100%

101%

0.02%

0.09%

0.16%

100%

103%

CE2-3.1b

0.13%

0.14%

0.20%

100%

99%

0.17%

0.16%

0.26%

100%

100%

CE2-3.1c

0.03%

0.12%

0.11%

100%

99%

0.03%

0.09%

0.18%

100%

99%

CE2-3.1d

0.22%

0.34%

0.42%

100%

99%

0.07%

0.15%

0.30%

100%

99%

Animation

CE2-3.1a

-0.02%

-0.01%

-0.04%

100%

103%

0.02%

0.06%

0.02%

100%

99%

CE2-3.1b

0.04%

0.02%

0.01%

101%

103%

0.03%

0.02%

-0.01%

101%

100%

CE2-3.1c

0.10%

0.20%

0.18%

100%

102%

0.09%

0.21%

0.30%

101%

99%

CE2-3.1d

0.21%

0.35%

0.38%

100%

100%

0.10%

0.14%

0.19%

101%

99%

Mixed Content

CE2-3.1a

0.08%

0.01%

0.00%

100%

102%

0.02%

0.07%

0.11%

100%

100%

CE2-3.1b

0.33%

0.25%

0.21%

101%

101%

0.28%

0.29%

0.15%

101%

100%

CE2-3.1c

0.06%

0.03%

0.06%

100%

102%

0.00%

0.01%

0.00%

100%

100%

CE2-3.1d

0.13%

0.23%

0.20%

100%

100%

0.22%

0.27%

0.24%

101%

100%

Camera-Captured

CE2-3.1a

0.00%

0.00%

0.00%

99%

96%

0.00%

0.00%

0.00%

101%

97%

CE2-3.1b

0.00%

0.00%

0.00%

101%

98%

0.00%

0.00%

0.00%

101%

98%

CE2-3.1c

0.00%

0.00%

0.00%

100%

97%

0.00%

0.00%

0.00%

101%

97%

CE2-3.1d

0.00%

0.00%

0.00%

101%

96%

0.00%

0.00%

0.00%

101%

98%

The current design requires checking maximum 64 predictors regardless of the block size. Worst case would be 4x4 blocks with this regard.

  • Method a restricts the number of checks to 16 for 4x4, 32, for 4x8/8x4, 64 for >=64 samples.
  • Methods b is even more restrictive, but also has slightly more loss
  • Methods c and d completely inhibit the update for small blocks

Method a is the most appropriate solution in terms of complexity benefit vs. compression.

It is, however, not obvious that the latency problem in palette mode is really severe, compared to other elements of VVC with 4x4 block size (in particular, intra and IBC).

No need for action was identified on this.

Decisions
No need for action was identified on this.
Citation