JVET-AD0401 BoG Report on Metrics for Tool Complexity Assessment [X. Li]

This document contains the report of the BoG on metrics for tool complexity assessment. The BoG met on 4/27 Thursday 6:10pm – 7:30pm.

To discuss how to proceed on defining better metrics (In the VVC development, there were other criteria such as number of worst-case memory accesses, number of operations, number of cycles before a certain value would be available, etc) for tool complexity assessment.

• Recommendations: Increase the precision of runtime ratio by 1 decimal digit in the ECM result template
• Add a mandate to AHG7 to develop methodology of more reliable runtime measurement
• Include efficiency vs runtime figures (separated figures for encoding time and decoding time) in AHG7 report
• Report total memory usage numbers for RA and LDB/P (by the JVET-Z0150 output in encoding log) in AHG7 report and evaluate the variation of the numbers. If the numbers are stable, further suggest including the memory usage numbers in ECM result template

This was presented Friday 28 April at 0900.

No specific discussion was necessary, and this was to be implemented in AHG work. It requires modification of the Excel reporting template.

AHG8: Optimization of encoders and receiving systems for machine analysis of coded video content (10)

Contributions in this area were discussed at 1600–2010 on Monday 24 April 2023 (chaired by JRO).

JVET-AD0401 BoG Report on Metrics for Tool Complexity Assessment [X. Li]

See section 4.9 for notes on this BoG report.

Liaison communications (1)

m63436 Liaison statement from SC 29/WG 1 to WG 5 on JPEG AI [WG 1 via SC 29 Secretariat]

TD 168/Gen LS/r on JPEG AI (SG16-LS53) [from ISO/IEC JTC1/SC29/WG1]

Similar incoming liaison statements were sent to WG 5 and VCEG to exchange information on the progress of work on neural network image coding in the JPEG AI project, for coordination with the work in JVET on NNVC.

The JPEG AI Verification Model 1 had been released after the JPEG 97th meeting (the meeting before the January 2023 meeting). JPEG AI Verification Model (VM) 1 main targets the standard reconstruction task. The JPEG AI VM has the following characteristics: 1) Entropy decoding was decoupled with prediction of latents and reconstruction of samples done independently, 2) Auto-regressive latent sample reconstruction module was parallelized with wavefront processing, 3) Latents were predicted and only the residual was coded and transmitted, 4) Hyper scale decoder provided the estimation of the variance of the entropy coding model distribution, 5) Gain units provided rate adaptation.

There were two key configurations: "tools off" which includes minimum set of subnetworks for encoding and decoding, "tools on" which additionally includes NNbased elements improving compression performance with better content adaptation. JPEG AI VM1 was reported to demonstrate 28% (tools-off) / 31% (tools-on) compression gain over JPEG’s VVC anchor for camera-captured content. The JPEG AI VM decoder currently required 780 (tool-off), 890 (tools-on) kMAC/pxl.

A new JPEG AI test set was released during the January 2023 JPEG meeting. This dataset for the evaluation of the JPEG AI VM contained 50 images, which was reported to avoid overfitting and allow to track the performance improvements from meeting to meeting. The JPEG AI Common Training and Test Conditions were updated to include this new dataset.

In the January 2023 JPEG meeting, it was also decided to integrate several changes into the JPEG AI VM2, speeding up training, improving performance at high, fixing bugs. The JPEG AI VM Software Guidelines were approved, describing the initial setup repository of JPEG AI VM, how to obtain the JPEG AI dataset, how to run tests and training. A description of the structure of the JPEG AI VM repository was also available.

A set of core experiments were established at this meeting targeting RD performance and complexity improvements. More information about JPEG AI results, VM and core experiments was in attached output documents of JPEG.

A reply was drafted by JVET as WG 5 N 211. The draft reply was presented in the AG 3 meeting Thursday 10:00 (G. Sullivan) and was reviewed in JVET on Thursday 27 April 1740.

Project planning

Software timeline

ECM9 software (including all adoptions) was planned to be available 3 weeks after the meeting.

The NNVC 5.0 codebase software (integrating “low” operation point loop filter) was planned to be available 2 weeks after the meeting. An update 5.1 (also including verified “high” operation point) was planned to be available after training verification.

VTM21.0 software was planned to be available on 2023-05-30. (Note that further updates may be released later)

Updates on top of HM17.0 software were not planned, but might be released after merging pending requests, as appropriate.

Core experiment and exploration experiment planning

An EE on neural network-based video coding was established, as recorded in output document JVET-AC2023.

An EE on enhanced compression technology beyond VVC capability using techniques other than neural-network technology was also established, as recorded in output document JVET-AC2024.

Initial versions of these documents were presented and approved.

Drafting of specification text, encoder algorithm descriptions, and software

The following agreement has been established: the editorial team has the discretion to not integrate recorded adoptions for which the available text is grossly inadequate (and cannot be fixed with a reasonable degree of effort), if such a situation hypothetically arises. In such an event, the text would record the intent expressed by the committee without including a full integration of the available inadequate text.

Plans for improved efficiency and contribution consideration

The group considered it important to have the full design of proposals documented to enable proper study.

Adoptions need to be based on properly drafted working draft text (on normative elements) and HM/VTM encoder algorithm descriptions – relative to the existing drafts. Proposal contributions should also provide a software implementation (or at least such software should be made available for study and testing by other participants at the meeting, and software must be made available to cross-checkers in EEs).

Suggestions for future meetings included the following generally-supported principles:

• No review of normative contributions without draft specification text
• VTM algorithm description text is strongly encouraged for non-normative contributions
• Early upload deadline to enable substantial study prior to the meeting
• Using a clock timer to ensure efficient proposal presentations (5 min) and discussions

As general guidance, it was suggested to avoid usage of company names in document titles, software modules etc., and not to describe a technology by using a company name.

General issues for experiments

It was emphasized that those rules which had been set up or refined during the 12th JVET meeting should be observed. In particular, for some CEs of some previous meetings, results were available late, and some changes in the experimental setup had not been sufficiently discussed on the JVET reflector.

Group coordinated experiments have been planned as follows:

• “Core experiments” (CEs) are the coordinated experiments on coding tools which are deemed to be interesting but require more investigation and could potentially become part of a draft standard by the next meeting or in the near future.
• “Exploration experiments” (EEs) are also coordinated experiments. These are conducted on technology which is not foreseen to become part of a draft standard in the near future. The investigating methodology for assessment of such technology can also be an important part of an EE. (Further general rules for EEs, as far as deviating from the CE rules below, should be discussed in a future meeting. For the current meeting, procedures as described in the EE description document are deemed to be sufficient.)
• A CE is a test of a specific fully described technology in a specific agreed way. It is not a forum for thinking of new ideas (like an AHG). The CE coordinators are responsible for making sure that the CE description is complete and correct and has adequate detail. Reflector discussions about CE description clarity and other aspects of CE plans are encouraged.
• A description of each experiment is to be approved at the meeting at which the experiment plan is established. This should include the issues that were raised by other experts when the tool was presented, e.g., interference with other tools, contribution of different elements that are part of a package, etc. The experiment description document should provide the names of individual people, not just company names.
• Software for tools investigated in a CE will be provided in one or more separate branches of the software repository. Each CE will have a “fork” of the software, and within the CE there may be multiple branches established by the CE coordinator. The software coordinator will help coordinate the creation of these forks and branches and their naming. All JVET members will have read access to the CE software branches (using shared read-only credentials as described below).
• During the experiment, revisions of the experiment plans can be made, but not substantial changes to the proposed technology.
• The CE description must match the CE testing that is done. The CE description needs to be revised if there has been some change of plans.
• The CE summary report must describe any changes that were made in the process of finalizing the CE.
• By the next meeting it is expected that at least one independent cross-checker will report a detailed analysis of each proposed feature that has been tested and confirm that the implementation is correct. Commentary on the potential benefits and disadvantages of the proposed technology in cross-checking reports is highly encouraged. Having multiple cross-checking reports is also highly encouraged (especially if the cross-checking involves more than confirmation of correct test results). The reports of cross-checking activities may (and generally should) be integrated into the CE report rather than submitted as separate documents.
• It is mandatory to report encoder optimizations made for the benefit of a tool, and if an equivalent optimization could be applied on the anchor, a comparison against the improved anchor shall be provided.
• A new proposal can be included in a CE based on group decision, regardless if an independent party has already performed a cross-check in the meeting when it was first proposed.

It is possible to define sub-experiments within particular CEs, for example designated as CEX.a, CEX.b, etc., where X is the basic CE number.

As a general rule, it was agreed that each CE should be run under the same testing conditions using one software codebase, which should be based on the group test model software codebase. An experiment is not to be established as a CE unless there is access given to the participants in (any part of) the CE to the software used to perform the experiments.

The general agreed common conditions for single-layer coding efficiency experiments for SDR video are described in the prior output document JVET-T2010.

Experiment descriptions should be written in a way such that it is understood as a JVET output document (written from an objective “third party perspective”, not a proponent perspective – e.g. not referring to methods as “improved”, “optimized”, “enhanced”, etc.). The experiment descriptions should generally not express opinions or suggest conclusions – rather, they should just describe what technology will be tested, how it will be tested, who will participate, etc. Responsibilities for contributions to CE work should identify individuals in addition to company names.

CE descriptions contain a basic description of the technology under test, but should not contain excessively verbose descriptions of a technology (at least not unless the technology is not adequately documented elsewhere). Instead, the CE descriptions should refer to the relevant proposal contributions for any necessary further detail. However, the complete detail of what technology will be tested must be available – either in the CE description itself or in documents that are referenced in the CE description that are also available in the JVET document archive.

Any technology must have at least one cross-check partner to establish a CE – a single proponent is not enough. It is highly desirable have more than just one proponent and one cross-checker.

The CE development workflow is described at:

https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/wikis/Core-experiment-development-workflow

CE read access is available using shared accounts: One account exists for MPEG members, which uses the usual MPEG account data. A second account exists for VCEG members with account information available in the TIES informal ftp area (IFA) system at:

https://www.itu.int/ifa/t/2017/sg16/exchange/wp3/q06/vceg_account.txt

Some agreements relating to CE activities were established as follows:

• Only qualified JVET members can participate in a CE.
• Participation in a CE is possible without a commitment of submitting an input document to the next meeting. Participation is requested by contacting the CE coordinator.
• All software, results, and documents produced in the CE should be announced and made available to JVET in a timely manner.
• A JVET CE reflector will be established and announced on the main JVET reflector. Discussion of logistics arrangements, exchange of data, minor refinement of the test plans, and preparation of documents shall be conducted on the JVET CE reflector, with subject lines prefixed by “[CEx: ]”, where “x” is the number of the CE. All substantial communications about a CE other than such details shall take place on main JVET reflector. In the case that large amounts of data are to be distributed, it is recommended to send a link to the data rather than the data itself, or upload the data as an input contribution to the next meeting.

General timeline for CEs

T1= 3 weeks after the JVET meeting: To revise the CE description and refine questions to be answered. Questions should be discussed and agreed on JVET reflector. Any changes of planned tests after this time need to be announced and discussed on the JVET reflector. Initially assigned description numbers shall not be changed later. If a test is skipped, it is to be marked as “withdrawn”.

T2 = Test model software release + 2 weeks: Integration of all tools into a separate CE branch of the VTM is completed and announced to JVET reflector.

• Initial study by cross-checkers can begin.
• Proponents may continue to modify the software in this branch until T3.
• 3rd parties are encouraged to study and make contributions to the next meeting with proposed changes

T3: 3 weeks before the next JVET meeting or T2 + 1 week, whichever is later: Any changes to the CE test branches of the software must be frozen, so the cross-checkers can know exactly what they are cross-checking. A software version tag should be created at this time. The name of the cross-checkers and list of specific tests for each tool under study in the CE plan description shall be documented in an updated CE description by this time.

T4: Regular document deadline minus 1 week: CE contribution documents including specification text and complete test results shall be uploaded to the JVET document repository (particularly for proposals targeting to be promoted to the draft standard at the next meeting).

The CE summary reports shall be available by the regular contribution deadline. This shall include documentation about crosscheck of software, matching of CE description and confirmation of the appropriateness of the text change, as well as sufficient crosscheck results to create evidence about correctness (crosscheckers must send this information to the CE coordinator at least 3 days ahead of the document deadline). Furthermore, any deviations from the timelines above shall be documented. The numbers used in the summary report shall not be changed relative to the description document.

CE reports may contain additional information about tests of straightforward combinations of the identified technologies. Such supplemental testing needs to be clearly identified in the report if it was not part of the CE plan.

New branches may be created which combine two or more tools included in the CE document or the VTM (as applicable).

It is not necessary to formally name cross-checkers in the initial version of the CE description document. To adopt a proposed feature at the next meeting, JVET would like to see comprehensive cross-checking done, with analysis of whether the description matches the software, and a recommendation of the value of the tool and given tradeoffs.

The establishment of a CE does not indicate that a proposed technology is mature for adoption or that the testing conducted in the CE is fully adequate for assessing the merits of the technology, and a favourable outcome of CE does not indicate a need for adoption of the technology into a standard or test model.

Availability of specification text is important to have a detailed understanding of the technology and also to judge what its impact on the complexity of the specification will be. There must also be sufficient time to study this in detail. CE contributions without sufficiently mature draft specification text in the CE input document should not be considered for adoption.

Lists of participants in CE documents should be pruned to include only the active participants. Read access to software will be available to all members.

Establishment of ad hoc groups

The ad hoc groups established to progress work on particular subject areas until the next meeting are described in the table below. The discussion list for all of these ad hoc groups was agreed to be the main JVET reflector (jvet@lists.rwth-aachen.de).

Review of AHG plans was conducted during the plenary on Friday 28 April 2023 at 1000.

It was confirmed that the rules which can be found in document ISO/IEC JTC 1/‌SC 29/‌AG 2 N 046 “Ad hoc group rules for MPEG AGs and WGs” (available at https://www.mpegstandards.org/adhoc/), are consistent with the operation mode of JVET AHGs. It is pointed out that JVET does not maintain separate AHG reflectors, such that any JVET member is implicitly a member of any AHG. This shall be mentioned in the related WG Recommendations. The list above was also issued as a separate WG 5 document (ISO/IEC JTC 1/‌SC 29/‌WG 5 N 212) in order to make it easy to reference.

Output documents

The following documents were agreed to be produced or endorsed as outputs of the meeting. Names recorded below indicate the editors responsible for the document production. Where applicable, dates of planned finalization and corresponding parent-body document numbers are also noted.

It was reminded that in cases where the JVET document is also made available as a WG 5 output document, a separate version under the WG 5 document header should be generated. This version should be sent to GJS and JRO for upload.

The list of JVET ad hoc groups was also issued as a WG 5 output document WG 5 N 212, as noted in section 9.