JVET-Q0826 Suggested content for summary information on BD-rate experiment evaluation practices [J. Ström, K. Andersson, R. Sjöberg (Ericsson), F. Bossen, A. Segall (Sharp), G. J. Sullivan (Microsoft), J.-R. Ohm (RWTH Aachen)]

This document provides general summary information about typical quality measurement practices for video coding experiments in recent work in MPEG and VCEG and their JVET and JCT-VC joint teams. In particular, it describes the use of Bjøntegaard Delta rate (BD-rate) measurements. It aims at providing a concept-level overview of recent practices and to provide some references to technical papers that describe further details. It provides comments on why some of the choices were made, and points out situations where caution must be taken when interpreting the results.

It is suggested that JVET use this document as the basis of an output document providing such information to the community.

Discussed aspects

• Add comment on averaging when some frame(s) are near lossless; this can cause a single very high PSNR value.
• Clarify comment on division by zero
• Log domain for the curve fitting, add figure for the log domain
• 6:1:1 vs. 8:1:1
  • AHG13 averages BD rates, not PSNRs
• Extrapolation (interp only, there is no extrap)
• Clarify integration and divide by log rate range to obtain an average
• Comparing piecewise cubic with cubic fit to check validity
• 37.5394 🡪 37.54
• wPSNR for 360° 🡪 WSPSNR
• Error in citation [5]

Withdrawn (0)

Section kept for future use.

Plenary meetings, joint meetings, BoG reports, and summary of actions taken

High-level syntax / systems relation meeting

This planned session was cancelled due to a lack of identified need.

Plenary meeting Sunday 12 Jan. 0800-1215

Reports of the tracks were presented as follows:

The status of Tracks A and B was presented and discussed, which particularly included the following aspects:

Track A:

• CE4 and CE4-related (some addressed in BoG): Geometric prediction mode (0.3% RA, 0.7% LB, relative to the non-triangle anchor with approx. same complexity 0.7% RA, 1.6% LB)
  • CE4 test of harmonized design removing separate triangle mode confirmed: coding efficiency benefit, likely subjective benefit and possible other utility, no significant complexity impact (although extra verification effort), so mode adopted
  • Simplification by reducing number of partition modes to 64 from 82
  • Further simplification / consistency cleanup by sampling position change JVET-Q0188
  • Further complexity reduction & text simplification by LUT replacement with equation JVET-Q0365
  • Further complexity / verification effort reduction by disabling for 8x64 and 64x8
• Open question: Syntax level for signalling number of merge candidates (for multiple uses)
  • This involves four syntax elements –for regular mode, for triangle (geometry), for IBC and for subblock. They should be handled consistently, at the same syntax level. It was agreed (Saturday in Track A CE4 related discussion) that as far as the triangle/geometry control perspective is concerned, SPS level control over the number of merge candidates seemed sufficient. As far as Track A was concerned, establishing the max number of merge candidates for all of these at the SPS level without lower-level control seemed sufficient.

In the plenary, it was agreed to move all this control to the SPS level. There was a suggestion of a flag and conditionally sending a number minus 1 (except for geometry mode, which is to be signalled as a delta similar to VVC draft 7). This was agreed to later be further discussed in Track A for review of candidate text provided by Ling Li, Y.-J. Chang and S. Esenlik.

• High-level syntax
  • Various minor cleanups, bug fixes, clarifications, highlights:
    • Restructuring of the PH
    • POC LSB in the PH rather than the SH
    • Considering to allow PH to be put into SH for single-slice use
  • Two constraints similar to AVC and HEVC profiles
    • Disallow separate colour plane mode in the 4:4:4 profile(s); it was agreed to further discuss this in Track A and check whether there is a functioning specification of that case, consider renaming or removing the flag (and no action was ultimately taken on it).
    • Disallow in base and dependent enhancement having layer differences of bit depth and chroma format (different chroma positions allowed)
  • A substantial amount of not-yet-presented contributions and further discussion topics remained, which was resolved later.
  • A major BoG that had not yet finished or reported to the track

For software coordination, there was discussion of whether to set up a separate reflector for that, but it was agreed that such discussions can just be conducted on the main JVET reflector.

It was agreed in the plenary that the palette mode is to be prohibited in the planned 4:2:0 profiles.

Track B:

CE1: Deblocking

• Subjective (informal) viewing conducted
• Adoption of JVET-Q0054, variant CE1-1-1. This resolves oversmoothing of long deblock filter.
• No action on other sub-CEs

CE2: Palette. No action from the CE itself, but there is a contribution related to CE2-1 recommended for adoption (improvement at low QP and lossless)

CE3: Lossless and near lossless

• Adoption of JVET-Q0088, CE3-2.3, high-level switch between RRC and TSRC for TS.
• Further candidate for adoption: BDPCM for chroma in 4:2:0 (JVET-Q0089), extension on CE3-2.3 into CE3-2.4. Pending further discussion on availability of CTC results
• This makes VVC slightly more efficient for lossless than HEVC, but more likely application of these modes is mixed lossy/lossless
• TSRC turns out rather being an SCC tool, but is efficient for SC.

CE5: CCALF

• Subjective (informal) viewing performed, visual artefact problem of CCALF resolved by encoder modification
• Candidate for adoption: CE5-2.1 (JVET-Q0073), simplification of common base, replacing multiplications by shift operations, same filter shape
• Roughly 1% gain in AI, 1.2-1.3% in RA, 1% in LB/LP (when shifting rate from chroma to luma)
• Pending further review of CE5 related contributions, looking for possible further improvements of CE5-2.1
• Desire to enable/disable at SPS, following ALF, no operation separate from ALF

Non-CE:

• Deblocking (CE1 related): Nothing noteworthy
• Loop filtering: Some cleanups, contribution on virtual ALF boundary problem (subjective viewing)
• Palette (CE2 related): Some cleanups and bug fixes, predictor initialization in WPP (JVET-Q0501), improved escape coding at low rates (JVET-Q0491), palette for non-4:4:4 (JVET-Q0503) – not intended for 4:2:0 profile, but 4:2:0 and 4:2:2 capable, restriction of 4x4 blocks
• Lossless coding (CE3 related): Nothing noteworthy
• Intra prediction: Some cleanups and bug fixes, MIP with constant shift/offset without loss (JVET-Q0446)
• Reference picture resampling: Only low-level reviewed, bug fix filters for normal blocks (JVET-Q0449) and downsampling filters for affine blocks (JVET-Q0517) -> most remaining aspects HLS, for Track A
• Transforms and signalling: Some cleanups and bug fixes on MTS and LFNST
• Entropy coding: Update zero word threshold, only for high tier (JVET-Q0436)
• Quantization: Bugfix on inconsistency chroma QP adaptation, other items ongoing
• Remaining items: Partitioning, ACT and other chroma tools, CE5 related (BoG), inter prediction (new in Track B, BoG), other tools, encoder optimization, etc.

Decisions recommended from trackA and B were agreed and approved, unless otherwise noted:

RPR HLS aspects were discussed in the plenary and were agreed to be considered in Track A.

Subjective viewing for VTM vs. HM was encouraged to be planned if feasible. M. Wien was requested to coordinate this.

Mathias, Ken, Dan and some other experts will identify suitable rate/quality points for VVC vs. HEVC comparison, in coordination with Vittorio, such that later during the week “dry run” tests could be done for checking the selection.

The number of contributions remaining to be considered in the following areas were noted:

• Text and software development (2 → 1 )
• Test conditions (0)
• Performance assessment (3)
• Coding studies and tools on specific use cases (3 → 4 → 3) – Film grain was discussed in the plenary and planned for further study in an AHG.
• Test Material (0)
• Conformance (1)
• Implementation studies (11) – moved most to inter prediction
• Profile/level specification (4) – two for parent-body joint meeting discussion, two others more specific

Among implementation studies, two software decoder implementation contributions were presented.

Conformance testing was discussed (see section 4.6).

Profile, tier and level specifications were discussed (see section 4.6).

Joint meeting Tuesday 14 January 0900-1000

A joint meeting with VCEG and MPEG Requirements was held at 0900 on Tuesday in Room 100 to discuss requirements for VVC (particularly JVET-Q0111 and JVET-Q0485).