International organisation for standardisation organisation internationale de normalisation


Notes from the BoG (starting at 17:19h)



Yüklə 8,2 Mb.
səhifə72/277
tarix02.01.2022
ölçüsü8,2 Mb.
#13030
1   ...   68   69   70   71   72   73   74   75   ...   277
Notes from the BoG (starting at 17:19h)

The question was raised whether open GOP or closed GOP is used in the IVC RA streams. The proponent asserts that the random access points are IDRs like IDRs in AVC. This corresponds to a closed GOP structure.

For VCB, the latest encoder dumps provided by the proponent had been partially assessed by an expert. The quantizer relation was checked and no deviations from the expected pattern were noted.

No further comments were raised regarding the AVC and WebVC bitstreams. It was stated that verification had been performed among contributing companies before releasing the bitstreams.

Follow-up discussion were held Wed. at 12:00 (chaired by J. Ohm)

The First round of “expert viewing” (with three non-video-experts) did not provide meaningful results. Therefore, it was decided to hold another round of viewing Wed. afternoon with a total of 12 video experts.

It is expected that with a sufficient number of experts, systematic outliers would be detected, and therefore trials to vote for one's “own codec” would result in invalid votes.

The question was raised whether it is possible to include the votes from the first round. It was agreed to leave to the test chair (Vittorio Baroncini) to decide whether this would be meaningful.

There was discussion about the suggested ranking method which was proposed by V. Baroncini and M. Wien – which was as follows: Compute, for each codec in each test case, the following credit number: 1+(count_of_worse-performing codecs). (Options of credit numbers: 1, 2, 5, 10 were discussed.) Compute an average over the different test cases.

It could (hopefully) be expected that the HP anchors usually have the highest score, this would give more credits to those codecs which might beat the anchor in some cases.

It was agreed to use this as indicator.

Question: What happens if two proposals are equal?



  • This has the effect that two equal proposals get one score lower (and if everything is undistinguishable, everybody gets close to nothing) – it was agreed that this is reasonable.

  • Would there be a threshold of counting “equal performance”? Likely to be based on confidence intervals – leave to test coordinators to decide whether CI can be determined reasonably.

Results were asked to be reported Thu. 9:00 in the Video room

Additionally, an SNR comparison was to be presented, with the following evaluation which could potentially give more evidence



  1. Applying the same ranking method as above, using data per rate point/sequence.

  2. Bjøntegaard values in 4 versions, where each of the four codecs is once used as anchor.

Presentation of results was planned for Thursday 9:00 (in a session chaired by J. Ohm):

Additional expert viewing was performed Wednesday afternoon (with 12 experts participating).

All subjects viewed all sequences, and confidence intervals were computed (V. Baroncini believed this was appropriate, though only 12 votes were used, since the votes were rather consistent).

BQ mall could not be used (as the anchor was corrupted).

720p testing was not possible to complete due to lack of time.

Evaluation with the ranking method was to be done, taking into account confidence intervals (to be presented in video plenary).

The tendency from the preliminary results of the visual test indicated the following:

For CS1: Random access (as tested, without 720p):



  • HP profile anchor (as tested) performed best

  • Other proposals (as tested) performed similarly to each other, with some tendency of VCB to perform slightly better than WVC/IVC

For CS2: Low delay (as tested, without 720p):

  • VCB (as tested) performed similar or slightly better than the HP anchor

  • WVC (as tested) performed similar or slightly worse than the HP anchor

  • IVC (as tested) was consistently worse than the others

It was mentioned by the IVC contributors that for the LD case, ITM7 performed worse than ITM6, which may be due to a bug, and badly influences CS2 results.

PSNR based BD rate averages showed >30% rate increase compared to AVC HP anchors for RA with all three other codecs; compared against each other rather than to the AVC HP anchors, the BD rate differences were marginal. For the LD case VCB had approx. 10%, WVC approx. 20%, IVC approx. 30% higher rate than anchor; also compared against each other, VCB performed obviously better in LD than both other codecs based on BD rate criteria.

For 720p (which was not tested visually) WVC and VCB were equivalent in quality according to PSNR based BD rate, in the LD case.

Generally, comparing the three codecs against each other (as tested), mutual comparison of PSNR by overall tendency confirmed the results of the visual test.

In previous meetings, doubts had been raised about the validity of the conclusion from the recent CfP submission concerning VCB, because


  1. The rate points did not match in many cases

  2. The influence of rate control was not clear

The new informal visual tests appeared to indicate

  1. That VCB fulfils quality expectations visually compared to other codecs with closely matching rates

  2. That rate control (which was not used here) was not an important factor in terms of the visual quality of VCB overall

It was, however, noted that without the rate control, it was not possible to hit the intended lowest bit rate point for the Party Scene sequence, and further clarification is needed whether this is a flaw in the normative design or can be solved by further modified encoder design.

For IVC, informal viewing tests had been performed by the last meeting, however the results by that time were incomplete since only few sequences had been found where the bit rates did closely match with the rates of the anchors (which were AVC HP for RA and AVC CBP for LD case), and only a few people had participated in the test.

The new results are more complete in this regard. For the RA case, the results of the last meeting are somewhat confirmed in that IVC performs still lower than AVC HP, but additionally it is found here that for the RA case IVC has a comparable quality as WVC and VCB.

For the low delay case, a consistently worse performance of IVC was found relative to the anchors and the other codecs, and the results of the last meeting are not confirmed. This may be due to some bug in the ITM7 implementation relative to ITM6, and requires further investigation.

Comparing the three “type 1” projects under consideration,


  • VCB and WVC are in a similar range of compression performance in the RA cases, whereas for LD, VCB has better performance (which may be due to a better optimized encoder)

  • IVC needs more improvement for the low delay case. More investigation is necessary to make it technically mature.

It is preliminary concluded that VCB (from the compression performance point of view) has been demonstrated to be a suitable candidate for standardization. However, doubts were raised regarding the maturity of the technical description. A new version of the VCB draft was to be uploaded and to be reviewed in the Video plenary Thu 1600 – potentially continuing Fri 08:00.

The following documents were all discussed in the VCC AHG and BoG (see above).

14.1.97.1.1.1.1.1.46m32268 AHG on VCC: AVC HP anchors [M. Wien (RWTH), K. Andersson, R. Sjöberg (Ericsson)]
14.1.97.1.1.1.1.1.47m32465 Report on the generation of the Web VC sequences [Kenneth R. Andersson, Krasimir Kolarov, Pierrick Philippe, David Singer, Rickard Sjöberg, Alexis Michael Tourapis]
14.1.97.1.1.1.1.1.48m32471 Submission of VCB streams to VCC AHG expert viewing [Mohamad Raad, Harald Alvestrand, Adrian Grange, Qunshan Gu]
14.1.97.1.1.1.1.1.49m32472 Updated MPEG-4 AVC HP anchors [Kenneth R. Andersson, Krasimir Kolarov, Pierrick Philippe, David Singer, Rickard Sjöberg, Alexis Michael Tourapis]
14.1.97.1.1.1.1.1.50m32476 Bug Fixes and Extensions to the MPEG-4 AVC JM Reference Software [Kenneth R. Andersson, Yuwen He, Krasimir Kolarov, Pierrick Philippe, David Singer, Rickard Sjöberg, Alexis Michael Tourapis]
14.1.97.1.1.1.1.1.51m32589 IVC bitstreams [Xufeng Li, Ronggang Wang, Siwei Ma, Tiejun Huang, Wen Gao]


  1. Yüklə 8,2 Mb.

    Dostları ilə paylaş:
1   ...   68   69   70   71   72   73   74   75   ...   277




Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur ©muhaz.org 2024
rəhbərliyinə müraciət

gir | qeydiyyatdan keç
    Ana səhifə


yükləyin