International organisation for standardisation organisation internationale de normalisation

4.6CE6: Intra prediction improvements

4.6.1Summary

4.6.1.1.1.1.1.1.1JCTVC-H0036 CE6: Summary Report of Core Experiments on Intra Coding Improvements [A. Tabatabai, K. Chono, E. Francois (CE Coordinators)]

The experiments are divided into 3 subsets, as listed below:

1. 
   Subset CE6a: Intra Chroma Prediction
   
2. 
   Subset CE6b: Improved Intra Prediction
   
3. 
   Subset CE6c: Intra mode coding cleanup and simplification
   

4.6.2Contributions

4.6.2.1.1.1.1.1.1Subtest a

Every proposal in this category adds some complexity:

• 
  G173 = H0295 is using reconstructed Cb residual as a refinement for Cr prediction.
  
• 
  G244 is checking irregularities to find the most suitable value for alpha in LM mode.
  
• 
  G346 uses reconstructed luma residual for chroma prediction additionally to LM mode.
  
• 
  G358 adds two new variants of LM mode (LML and LMA).
  

Note: The current HM5 has 6 chroma prediction modes, and G244, G346 and G358 introduce more modes.

(Document numbers of the previous meeting are used above rather than document numbers of this meeting, as these techniques had been proposed at the preceding meeting and had been discussed using those document numbers.)

None of the proposals adds a big advantage in compression performance.

Some concern was expressed that the intra prediction is already fairly complex (particularly in LM mode). H0295 seems to be the least complex, but introduces a new dependency in all modes. It was suggested to check the impact on hardware implementation.

For some architectures, there may be an increase in complexity (mostly at the encoder side, less for the decoder side).

One expert would be OK with having such a mode if it can be disabled by an SPS flag (which seems like not something we should do for a very specific prediction mode).

No action was taken in this area.
4.6.2.1.1.1.1.1.2JCTVC-H0117 CE6.a: Chroma intra prediction based on residual luma samples [K. Kawamura, T. Yoshino, H. Kato, S. Naito (KDDI)]
4.6.2.1.1.1.1.1.3JCTVC-H0118 CE6.a: Cross-check results of intra chroma prediction [K. Kawamura, S. Naito (KDDI)]
4.6.2.1.1.1.1.1.4JCTVC-H0129 CE6.a: Cross-check results of intra chroma prediction [S. Matsuo, M. Matsumura, S. Takamura, H. Jozawa (NTT)]
4.6.2.1.1.1.1.1.5JCTVC-H0171 CE6: Results on experiments related to section a on Intra Chroma Prediction [C.Gisquet, E. François (Canon), Y.-J. Chiu, Y. Han (Intel), X. Zhang, O. Au (HKUST), K. Kawamura, T. Yoshino (KDDI)]
4.6.2.1.1.1.1.1.6JCTVC-H0172 CE6: Cross verification of results related to section CE6.A [E. François, C. Gisquet (Canon)]
4.6.2.1.1.1.1.1.7JCTVC-H0295 CE6.a: Cross-channel prediction refinement to improve intra chroma prediction [Y.J. Chiu, Y. Han, L. Xu, W. Zhang, H. Jiang (Intel)]

This contribution reports test results of the cross-channel prediction refinement technique to improve intra chroma prediction of HM 5.0. Presented at JCT-VC meeting at Geneva, the cross-channel prediction refinement technique produces a refinement signal based on a fixed weighting of the reconstructed cross-channel chroma residual signal. Two test data points of the chroma cross-channel prediction refinement are reported: global weighting prediction and mode-dependent weighting prediction. Compared to the HM5.0 anchor of the common test condition, the mode-dependent weighting prediction provides a reported average BD bit rate change of -0.3% on Y, -0.6% on Cb, and -3.4% on Cr on the AI HE test cases A-E, of -0.3% on Y, -0.8% on Cb, and -3.3% on Cr on AI LC test cases A-E, and of -0.3% on Y, -0.6% on Cb, and -3.3% on Cr on AI HE-10 test cases A-F, without significant impact on the CPU run time (1% decrease in Enc time, 0% increase in Dec time for AI HE case, 0% increase in Enc time, 1% increase in Dec time for AI LC case, and 1% decrease in Enc time, 0% increase in Dec time for AI HE-10 case).

4.6.2.1.1.1.1.1.8JCTVC-H0573 Cross Check for Intel intra chroma prediction (JCTVC-H0295) [K. Kawamura (KDDI)] [late]
4.6.2.1.1.1.1.1.9JCTVC-H0296 CE6.a: Cross-check report for CE6a on intra chroma prediction [Y.J. Chiu, Y. Han, L. Xu, W. Zhang(Intel)]
4.6.2.1.1.1.1.1.10JCTVC-H0323 CE6.a: Cross-check of three combined cases of the four intra chroma prediction tools by Huawei [H. Yang (Huawei)]
4.6.2.1.1.1.1.1.11JCTVC-H0539 CE6.a: Cross-check report for intra chroma prediction [X. Zhang, O. Au (HKUST)]
4.6.2.1.1.1.1.1.12JCTVC-H0544 CE6.a: New modes (LML and LMA) for chroma intra prediction [X. Zhang, O. Au, J. Dai, X. Wen, C. Pang, F. Zou (HKUST)]
4.6.2.1.1.1.1.1.13Subtest b

JCTVC-G135/H0437 proposes to add two prediction unit (PU) types (i.e. 2NxN and Nx2N) to Intra coding units (CU) in addition to the existing PU types (i.e. 2Nx2N and NxN) relative to the current HM design. It also reports NSQT harmonization results with and without 8x2/2x8 transforms.

JCTVC-G119/H0059 proposes to make the total number of remaining intra modes a power of two and use FLC encoding for the binarization.

It was asked what is the relationship between H0437 and H0455 (the latter being categorized as an AHG16 related contribution). It was suggested that they should give the same results, but are different. H0437 is subsumed in H0455 and it was suggested that H0437 can be discussed in the context of AHG16 discussions.

4.6.2.1.1.1.1.1.14JCTVC-H0437 CE6.b: Rectangular (2NxN and Nx2N) PU for Intra Prediction [X. Zhang, S. Liu, S. Lei (MediaTek)]
4.6.2.1.1.1.1.1.15JCTVC-H0297 CE6.b: Cross verification of MediaTek rectangular (2NxN and Nx2N) intra prediction (JCTVC-H0437) [Y.J. Chiu, Y. Han, L. Xu, W. Zhang (Intel)]
4.6.2.1.1.1.1.1.16JCTVC-H0346 CE6.b crosscheck for MediaTek proposal [Y. Lin, C. Lai, L. Liu, J. Zheng (HiSilicon)]
4.6.2.1.1.1.1.1.17JCTVC-H0268 CE6: Cross-check for Subtest CE6b of Rectangular (2NxN and Nx2N) Intra Prediction [S. Oudin (Fraunhofer HHI)]
4.6.2.1.1.1.1.1.18JCTVC-H0403 CE6.b: Cross-check report for Rectangular (2NxN and Nx2N) PU for Intra Prediction [J. Xu, A. Tabatabai (Sony)]
4.6.2.1.1.1.1.1.19JCTVC-H0187 CE6: Crosscheck of JCTVC-G119 [K Sato (Sony)]
4.6.2.1.1.1.1.1.20JCTVC-H0057 CE6 subset 5.2.2 and 6.2.2: Intra coding improvements [Y. Lin, L. Liu, J. Zheng (HiSilicon), H. Yang, H. Yu (Huawei)]

Relating to subtest b, this is removing the Hor+8 and Ver+8 modes and adding a new mode which is the average of both. In terms of compression, there was no benefit (at most 0.1% BR reduction). It reduces one mode, but needs to add in the averaging process, and therefore it is not a real simplification.

Relating to subtest c, removing the Hor+8 entirely resulted in no loss in compression.

4.6.2.1.1.1.1.1.21JCTVC-H0192 CE6: Cross-check report for Subtest CE6b and CE6c on modifications to Intra-frame coding [H. L. Tan, C. Yeo, Y. H. Tan (I2R)] [late]

4.6.2.1.1.1.1.1.22Subtest c

4.6.2.1.1.1.1.1.23JCTVC-H0075 CE6: Subtest 6c (6.2.1) – Intra mode coding simplification [C. Yeo, Y. H. Tan, H. L. Tan, Z. G. Li (I2R)]

This contribution proposed a modification to the intra prediction mode coding in the HM based on the proposal JCTVC-G153. The remaining mode coding uses a binary+FLC binarization. When the binary flag is coded in bypass mode, there is reportedly no loss in coding performance for AI-HE and AI-LC configurations. When the binary flag is coded using a context, there is reportedly an average BD bit rate gain of 0.2% for both AI-HE and AI-LC configurations.

General opinion: This is a straightforward approach of simplification.

It is conceptually similar to the 3 MPM approach

Various similar methods targeting simplification had been proposed (see CE summary) – similar to H0057 and H0075, and it was agreed to establish a breakout group on simplification of intra mode coding (coordinated by K. Chono) – see further discussion in section 5.17.

4.6.2.1.1.1.1.1.24JCTVC-H0058 Cross-verification of CE6 subset 6.2.1: On intra prediction mode coding (JCTVC-H0075) [Y. Lin, L. Liu (HiSilicon)]
4.6.2.1.1.1.1.1.25JCTVC-H0440 CE6.c: Cross-check of JCTVC-H0075: Intra mode coding simplification [X. Zhang, S. Liu (MediaTek)] [late]
4.6.2.1.1.1.1.1.26JCTVC-H0478 CE6.c: Cross check of Huawei/HiSilicon proposal JCTVC-H0057 [J Kim, C Kim, B Jeon (LG)]