Back to Search
Document details
Report of BoG on intra prediction and mode coding
Abstract
The BoG reviewed proposals related to the intra prediction and mode coding topics (CE3-related). This report includes the recommendations of the BoG for adopting or further studying of proposals. It is also suggested to revisit/present two MIP-related items in track A.
JVET-P0969 Report of BoG on intra prediction and mode coding [G. Van der Auwera]
Presented Monday 7 Oct. 1400-1615 in Track A (chaired by JRO)
The BoG reviewed proposals related to the intra prediction and mode coding topics (CE3-related). This report includes the recommendations of the BoG for adopting or further studying of proposals. It is also suggested to further discuss two MIP-related items in track A.
The BoG met during the following times:
- Friday 4 Oct.: 9:00-11:00, 14:00-19:45
- Saturday 5 Oct.: 11:00-13:00, 19:00-20:00
Summary of BoG recommendations and open issues:
- MIP:
- To be further discussed in track A:
- JVET-P0199, JVET-P0289, JVET-P0303: Problem addressed is that given a 4 × 16 or 16 × 4 MIP coded block, a subset of the matrix multiplication output with odd row indices is used, while for other MIP coded blocks all matrix rows are used. It is proposed to align the MIP matrix multiplication process by defining MipSizeId to be equal to 1 for a 4 × 16 or 16 × 4 MIP coded block.
Agreed in Track A that this is useful.
Decision: Adopt - JVET-P0803 proposes combination of CE3-3.2, aspect of JVET-P0194, JVET-P0198, JVET-P0352, JVET-P0398, JVET-P0535 (see detailed notes in the report):
- MIP upsampling order is fixed (adopted test CE3-3.2).
- (a) MIP is allowed up to 64x64 regardless of the maximum transform size. (Motivation: handle sizes MIP as normal intra, gain 0.02% in AI and RA)
- agreed to adopt this aspect. - (b) The following three together give 0.06% AI and 0.04% RA, Matrices were retrained as the previous matrices would not fit:
- Alignment that allows all MIP-modes to be transposed.
- The number of MIP modes is 32 for sizeId=0, 16 for sizeId=1 and 12 for sizeId=2. (Motivation: Cleanup after removal of MPM)
- Block aspect ratio restriction (<=4) is removed. (also allows long, narrow blocks)
- JVET-P0199, JVET-P0289, JVET-P0303: Problem addressed is that given a 4 × 16 or 16 × 4 MIP coded block, a subset of the matrix multiplication output with odd row indices is used, while for other MIP coded blocks all matrix rows are used. It is proposed to align the MIP matrix multiplication process by defining MipSizeId to be equal to 1 for a 4 × 16 or 16 × 4 MIP coded block.
- To be further discussed in track A:
During the discussion of track A, several experts expressed their opinion that this package is a reasonable cleanup, which removes several exception cases. Except for procedural concern (new proposal), no strong objection was raised, and additional information was provided as requested. Agreed to adopt this package of three items.
- (c) Separate context for the first bin of the MTS-index depending on MIP. (gain: 0.04% AI, 0.02% RA) – it was agreed that this is not a cleanup, and the gain is so low it is not justified.
Decision: Adopt Items (a) and (b). JVET-P0803 without the context modification.
The slide deck shown to be uploaded in a new version of JVET-P0803.
- For further study (some potentially in core experiment, if established):
- JVET-P0136: Reduces the MIP offset table (mode, sizeId). Can be studied in combination with CE3-2 test that simplifies the MIP right shifting operation. Suggested to start from floating point models to avoid double quantization.
- JVET-P0625: Proposes to remove MIP offset table and replace offset multiplication with a shift operation. Suggested to start from floating point models to avoid double quantization.
- JVET-P0509: Proposes to right shift the MIP input samples to 8-bit values before the matrix multiplication process so that MIP matrix multiplication process can be implemented with 8-bit SIMD multiplications. Suggested to further study of SIMD implementation and benefits. (From discussion in track A: Some concern was expressed that reducing the bit depth may have some impact on visual quality, and also may prohibit increased bit depth. Not so important to take action)
- JVET-P0560: Proposes to reduce the number of contexts for MIP mode flag, which is claimed to save line buffer storage for MIP flag. Suggested to further study in the context of potential changes to MIP this meeting. [From discussion in track A: Advantage not substantial enough to take action]
- For further study (some potentially in core experiment, if established):
The proponents of JVET-P0136, JVET-P0625 and CE3-2.1 are encouraged to perform further study on these proposals, based on the floating point models of the newly adopted matrices (to be obtained from J. Pfaff). The purpose of this study shall be removal of table lookups for better software implementation, in optimum case without changing the current behaviour in terms of bitstream output and performance. No CE on these aspects
- Intra mode coding:
- For adoption:
- JVET-P0615: Proposes to clean-up the sharing of context for cclm_mode_idx and intra_chroma_pred_mode syntax elements. It is more logical for each syntax element to have its own context. Decision: Adopt.
- For further study:
- JVET-P0369, JVET-P0536: Extend MRL with non-MPM intra modes. Suggested to further study the encoder search for non-MPM MRL modes. [From track A discussion: Further study on this is discouraged, unless it is shown to provide >0.15% benefit in terms of compression without substantial run time increase.]
- For adoption:
- Intra prediction – various topics:
- For adoption:
- JVET-P0111: Proposes cleanup to more accurately map angular modes from luma to chroma for 4:2:2 content (minor rounding changes to mode values in table). Agreed – Decision (BF): Adopt
- JVET-P0329: The proposal suggests one editorial change that removes nW = Max(nTbW,2) from the spec, which is not needed because width 1 partitions were removed last meeting (ISP simplification). The second suggestion is a simplification to also remove the nH = Max(nTbH,2) from the spec, which has the effect that planar is applied only one-dimensionally in case the block height is one.
Agreed – Decision (BF): Adopt - JVET-P0418: Proposes to modify the definition of MRL to use the same 3 lines as CCLM. It is claimed that it aligns MRL with CCLM in terms of neighbouring pixel referencing and reduces the storage requirements for decoders. 0.04% AI loss is reported. Agreed – Decision: Adopt
- JVET-P0599: Cleanup of interpolation filtering by replacing 4-tap “Gaussian” filter with the 4-tap smoothing filter obtained by convolving a linear filter with 3-tap FIR filter that has coefficients of [1, 2, 1]/4. The proposed filter coefficients are very similar to the current filter coefficients (fG in spec), but the observed frequency response issues with the current fG filter in VVC are fixed and some potential visual artefacts are claimed to be avoided. Agreed – Decision: Adopt
- For further study (potentially in core experiment, if established):
- JVET-P0500: It is proposed to extend CCLM with 4 additional modes (4 reference sample positions) in addition to LM, LM_A, LM_L (1.03%/1.06% U/V gains reported with negligible encoding/decoding time change). [The gain of around 1% only in chroma is fairly low and would not justify the additional modes (going from 3 to 7)]
- Editorial: JVET-P0626 - Decision: Left to the discretion of editor to change that.
- For adoption:
JVET-P0969 Report of BoG on intra prediction and mode coding [G. Van der Auwera]
See section 6.3.
References:
JVET-O0297
JVET-O0343
JVET-O0545
JVET-P0054
JVET-P0064
JVET-P0111
JVET-P0136
JVET-P0150
JVET-P0177
JVET-P0194
JVET-P0198
JVET-P0199
JVET-P0208
JVET-P0265
JVET-P0277
JVET-P0289
JVET-P0294
JVET-P0297
JVET-P0302
JVET-P0303
JVET-P0318
JVET-P0329
JVET-P0352
JVET-P0355
JVET-P0358
JVET-P0369
JVET-P0374
JVET-P0398
JVET-P0401
JVET-P0406
JVET-P0418
JVET-P0484
JVET-P0500
JVET-P0503
JVET-P0507
JVET-P0509
JVET-P0510
JVET-P0511
JVET-P0520
JVET-P0531
JVET-P0535
JVET-P0536
JVET-P0537
JVET-P0548
JVET-P0549
JVET-P0550
JVET-P0560
JVET-P0596
JVET-P0599
JVET-P0615
JVET-P0625
JVET-P0626
JVET-P0633
JVET-P0638
JVET-P0641
JVET-P0649
JVET-P0694
JVET-P0701
JVET-P0716
JVET-P0718
JVET-P0725
JVET-P0726
JVET-P0749
JVET-P0751
JVET-P0763
JVET-P0768
JVET-P0772
JVET-P0773
JVET-P0777
JVET-P0778
JVET-P0780
JVET-P0781
JVET-P0782
JVET-P0787
JVET-P0788
JVET-P0790
JVET-P0791
JVET-P0793
JVET-P0802
JVET-P0803
JVET-P0805
JVET-P0812
JVET-P0822
JVET-P0831
JVET-P0832
JVET-P0835
JVET-P0840
JVET-P0873
JVET-P0904
JVET-P0980
Decisions
Left to the discretion of editor to change that
See section 6.3.
Citation