International organisation for standardisation

6.2HM settings and common test conditions (5)

6.2.1.1.1.1.1.1.1JCTVC-F270 On BD-rate calculation [J. Wang, X. Yu, D. He (RIM)]

An issue about BD-rate calculation was observed and studied for Class-A sequences. This contribution summarizes the BD-rate issue and presents some recent studies that follow some suggestions from the reflector.

As Nebuta and BQTerrace are the most problematic sequences, it is recommended that BD-rate results excluding Nebuta and BQTerrace are also reported for evaluation purposes (and checked to see whether these are consistent with the behaviour for other data), and that additional information on these sequences such as the PNSR-bit rate curves are to be checked when inconsistencies are observed.

A BD-rate calculation based on piece-wise cubic interpolation reportedly appears to be a practical fix to this problem (COM 16 - C.404, April 2008, implemented in the included file hm32piecewisecubic2.xls, which originally came on the group email reflector from Frank Bossen).

It was commented that the implementation of the piece-wise cubic interpolation method might benefit from some refinement, and that for at least some time it may be desirable to have both results reported.

Having this (a spreadsheet that computes both metrics, using a somewhat better implementation of the piece-wise cubic method) seems like it would be a useful step forward to accomplish by shortly after this meeting. F. Bossen volunteered to provide that. Tentatively, it seems that the piece-wise cubic method seems somewhat more reliable.

Observing the curve shapes can reportedly also be helpful. Nebuta and BQTerrace seem to exhibit the problem partly due to being noisy sequences.

The amount of overlap between the curve ranges can also have unusual effects, and reporting the amount of overlap may be beneficial.

An issue about BD-rate calculation was observed and studied for Class-A sequences. This contribution summarizes the BD-rate issue and presents some recent studies that follow some suggestions from the reflector. As Nebuta and BQTerrace are the most problematic sequences, it is recommended that BD-rate results excluding Nebuta and BQTerrace are also reported for evaluation purpose and additional information on these sequences such as the PNSR-bitrate curves are to be checked when inconsistency is observed. The BD-rate calculation based on piece-wise cubic interpolation also appears to be a practical fix to this problem. It is recommended to further investigate the use of the current BD-rate tool in evaluation of proposals.

One suggestion that was made on the reflector is looking at BD values without the critical sequences, another would be to use the suggested piecewise cubic interpolation which appears to come closer to reality.

A *.xls file implementing the piecewise cubic interpolation is attached (implementation may not be perfect yet according to a comment made by F. Bossen).

A reasonable exercise would also be to investigate how different interpolation methods match with a densely sampled PSNR curve.

One possibility could be to compute BD values with old method and with the cubic interpolator and use divergence between both as an indicator for unreliable results for certain test cases.

Some experts express the opinion that inclusion of more test points may not solve the problem.

Frank Bossen will provide a spreadsheet implementing the cubic method in addition to the current method shortly after the meeting.

6.2.1.1.1.1.1.1.2JCTVC-F433 Reference Lists For Low Delay Settings [C. S. Lim, S. M. Thet Naing (Panasonic)]

This contribution provides an investigation report on construction of reference picture lists, and proposes a memory managing scheme for reference pictures under low delay constraints. The purpose of the investigation is to improve the coding efficiency of low delay settings by using a different set of reference pictures that includes more reference pictures that are of high quality. This contribution reports an average coding gain of 2.3% (HE) and 2.6% (LC) for low delay B settings and 2.2% (HE) and 2.3% (LC) for low delay P settings.

The suggestion was to use the two most recent high quality pictures plus the other two most recent (by temporal sequence) pictures, instead of four most recent pictures.

Does not increase the buffer requirements (in principle, but with current software implementation it would).

JCTVC-F493 proposes something very similar with similar gains.

Conclusion: Adopt this kind of change to LD settings in spirit.

Note: Current software uses hard-coded picture buffer management. In the future, this should be implemented by MMCO; JCTVC-F493 could be a solution on this.

A similar observation is reported in JCTVC-F493.

It was noted that this modification may (or may not) produce similar improvement in the JM context, which is desirable to study.

The group initially agreed, in principle, to adopt this non-normative modification to the common conditions. However, this adoption was later obsoleted following review of JCTVC-F519 and JCTVC-F701.

6.2.1.1.1.1.1.1.3JCTVC-F519 Cross-check of reference lists for low delay settings (JCTVC-F433) [B. Li (USTC), J. Xu (Microsoft)] [late upload 07-12]

It was reported that the proposed reference frame setting brought more than 2% bits saving on average under the default test configuration. It was reportedly clear that changing the reference frame setting for low delay case can bring significant bit saving. Similar investigations were also reportedly presented in JCTVC-F701, where the improvement capacity by adjusting reference frames for low delay cases is analyzed.

6.2.1.1.1.1.1.1.4JCTVC-F701 Encoding optimization to improve coding efficiency for low delay cases [J. Xu (Microsoft)] [late reg. 07-11, upload 07-12]

This document presents some encoder-only actions to improve the performance of the low delay configuration. Two methods are described: one is reference frame selection; the other is lambda and QP adjustment (bit allocation). The experimental results reportedly show that an average 9% bit rate saving can be obtained from that encoder only decision. The maximum bit rate savings is more than 19%.

Reference frame selection is similar as JCTVC-F433, here the nearest frame and one/three high-quality frame (in case of 2 /4 reference frames). Results for the case of 4 frames are slightly better than JCTVC-F433. Gain for 2 frames is higher.

Conclusion: For next LD settings, 3 high-quality frame plus the most recent additional frame shall be used (signalling still t.b.d.).

RDO optimization at sequence level – is this still low delay? Presumably not.

In principle, this means looking for frames that are more suitable as references.

It was remarked that this could also be done for RA – where this would be practical – this suggestion was encouraged to be investigated in further study.

6.2.1.1.1.1.1.1.5JCTVC-F696 Suggested SCC Test Conditions [X. Zhang, O. Au, C. Pang, X. Wen] [late reg. 07-10, upload 07-13]

This document proposed common test conditions and software reference configurations recommended to be used for the screen content coding experiments conducted between the 5th and the 6th JCT-VC meetings and between the 6th and the 7th JCT-VC meetings. These common test conditions were also proposed for use in technical contributions related to SCC for the 7th JCT-VC meeting.

Addressed by BoG and its results review.