International organisation for standardisation organisation internationale de normalisation

15.3Test model design: non-normative aspects and test methodology

15.3.1.1.1.1.1.1.1JCTVC-C253 Encoder speedup for bidirectional averaging with rounding control [R. Panchal, M. Karczewicz, P. Chen (Qualcomm)]

In the 2nd JCT-VC meeting, bidirectional averaging with rounding control (BiRound) was proposed and adopted in the TMuC. In this contribution further encoder optimizations were proposed leading to a reported 50% complexity reduction of the BiRound with no changes in the encoder bitstream.

In TMuC 0.8, a new function with 'BiRound' parameter is always used for bidirectional motion estimation. To achieve further speedup, the new function is used only when 'BiRound' parameter is 1. Otherwise, the original function is used.

Results were provided with LD HE configuration. The old version as implemented in TMuC 0.7 showed an increase in encoding time by 8% and 1.1% BR reduction, a speedup which produces bit-exact result is reported with 4% increase, and a sub-optimum speedup gives 0.9% reduction with no change.

Conclusion:

• 
  Decision: Adopted.
  
• 
  Should be on when IBDI is off and vice versa.
  
• 
  The fast bit-exact version shall be used in the encoder.
  

15.3.1.1.1.1.1.1.2JCTVC-C287 Cross-check of encoder speedup for bidirectional averaging with rounding control [B. Bross (Fraunhofer HHI)] (missing prior, available first day)

This contribution confirmed the bit rate results reported in JCTVC-C253, with the reported encoder speedup being slightly better.

15.3.1.1.1.1.1.1.3JCTVC-C315 On the computation of PSNR and BD-Rate [D. Hoang (Zenverge)] (late reg.)

This late contribution was not reviewed in group discussion due to lack of time.

The contribution asserted that the computations of PSNR and BD-Rate by the current TMuC software and spreadsheet macros were flawed, and recommendations were made on how to improve these computations.

16TMuC settings and common test conditions

16.1.1.1.1.1.1.1.1JCTVC-C084 Coding efficiency report of modification by TMuC draft005 [T. Yamakage, T. Chujoh, T. Watanabe (Toshiba)]

This contribution showed experimental results of the modification by TMuC draft005 that signals the ALF filtering control map in the slice header instead of signaling it in the CU header. Experimental results reportedly showed that the coding efficiency loss is 0.04% on average while the interleaving process of the ALF filtering control map at encoder side is removed.

This was one of the features of "TMuC low priority" disabled by default (only implemented after TMuC 0.7)

Several experts pointed out that this proposal simplifies the encoder (no buffering of encoded data).

Decision: There was no objection by the group, and it was agreed to Adopt this proposal.

One expert points out that there might be other ways to achieve this goal.

16.1.1.1.1.1.1.1.2JCTVC-C243 Cross check result of JCTVC-C084 [I. S. Chong, M. Karczewicz (Qualcomm)] (missing prior, uploaded substantially after the meeting)

This contribution was verbally presented prior to upload. It reportedly confirmed the results reported in JCTVC-C084.

16.1.1.1.1.1.1.1.3JCTVC-C155 TMuC 0.7 results with sequence dependent QPs [K. Andersson, R. Sjöberg (Ericsson)]

This contribution investigated performance when targeting CfP bit rates using sequence specific QPs rather than those specified in the common conditions (QP=22, 27, 32 and 37) for low delay and random access configurations. The sequence specific QPs was asserted to give very similar coding efficiency as for the common conditions and to give bit rates generally closer to application bit rates and thus to be better suited for subjective assessment. It was noted that the worst case encoding time was reduced by about 20% when using the proposed alternative values.

The following remarks were recorded in discussion of this contribution:

• 
  Typically, the example QP settings suggested have a much lower range of operational points – would this really be desirable?
  
• 
  Some cases go into the low-quality range particularly.
  
• 
  No action was taken on this, as the current settings seem appropriate for investigation of the merit of single coding tools.
  

16.1.1.1.1.1.1.1.4JCTVC-C168 Comments on common test conditions [K. Kondo, K. Sato, T. Suzuki (Sony)]

This contribution proposed to change the test conditions as follows.

• 
  To reduce the number of frames
  
• 
  To increase the number of test sequences
  

The common conditions test configurations were summarized in JCTVC-B300. Currently test sequences used by CfP are used and all frames must be encoded. Coding performance results with a smaller number of frames, e.g. 120, were reported to be not so different from the results of using all frames. Instead of reducing the number of total coded frames, it was proposed to test a larger variety of test sequences, e.g. VCEG test sequences, 4K sequence, 8K sequences, etc. To test a variety of test sequences was asserted to be more helpful to achieve an understanding of the the coding performance rather than encoding a larger number of frames in a smaller number sequences.

Comments recorded during the discussion were as follows:

• 
  Subjective tests should also be performed, such that long sequences are needed.
  
• 
  Extending the test set to provide more variety appears useful, but needs careful study.
  
• 
  Having long sequences is also an advantage, as it enforces to not use overly complex methods –with parallel processing, the length of the sequences is the main bottleneck anyway.
  
• 
  For extreme resolutions, a shorter length may be considered.
  
• 
  Work on fast motion estimation should be started.
  
• 
  It was suggested to conduct further study in an AHG (chaired by T. Suzuki)
  

17Application-specific topics

17.1.1.1.1.1.1.1.1JCTVC-C021 Evaluation result of JCTVC-B031 [K. Kazui, J. Koyama, A. Nakagawa (Fujitsu)]

A scheme which reportedly minimizes the loss of coding efficiency in very low delay coding was proposed at the 2nd JCT-VC meeting. This contribution shows evaluation result of the proposed scheme using TMuC software version 0.7.3.

Coding gains from the conventional multiple slices scheme in BD-Bit rate were reported as 9.0% and 15.9% when the sizes of LCTB are 64x64 and 32x32, respectively.

In addition, a new buffer model for a coding delay of less than one frame period was proposed in this contribution.

Remarks recorded during the discussion included the following:

• 
  A very high number of slices was used per frame – is this realistic? Certainly other solutions would be possible to achieve the same goal.
  
• 
  Constant bit rate was assumed for each slice – is that realistic?
  
• 
  It was suggested to contribute software to the work of AHG on slices.
  
• 
  Further study was encouraged.
  

17.1.1.1.1.1.1.1.2JCTVC-C032 Characteristics of Super Hi-Vision test sequences [Y. Shishikui, Y. Matsuo, A. Ichigaya, K. Iguchi, S. Sakaida (NHK)]

This contribution presented the characteristics of Super Hi-Vision (SHV) test sequences, which reportedly differ from those of the other test sequences. The analysis of test sequences indicated that SHV test sequences contain an extremely large amount of motion that affects spectrum characteristics of the signal to a large extent. SHV test signals also contain a larger noise component due to their high bandwidth, although this does not affect subjective picture quality as much as was expected by the objective amount.

The Sequences were interpolated by Bayer filter to achieve a full 8Kx4K resolution, and spectral characteristics are different from the 2K and 4K sequences (high flat-spectrum noise component appears to be present below quarter of sampling frequency).

17.1.1.1.1.1.1.1.3JCTVC-C055 Performance report of TMuC for Super Hi-Vision [K. Iguchi, A. Ichigaya, Y. Shishikui(NHK), S. Sekiguchi, A. Minezawa (Mitsubishi Electric)]

This contribution reported the performance of TMuC 0.7.0 for Super Hi-Vision (SHV) sequences. While the TMuC achieved substantial coding gain relative to the AVC CfP anchor, it took from 3.1 to 6.1 times encoding time than JM encoding with CfP anchor condition.

The following remarks and observations were recorded:

• 
  There was a suggestion that cropped sequences could be used that would give roughly same results with much lower simulation time.
  
• 
  BD bit rate improvement is approx. 60% for Steam Locomotive, 40% for Nebuta
  
• 
  PSNR curves show larger gap (higher gain) towards higher bit rates: Does the TMuC loop filter improve prediction under noisy conditions?
  
• 
  Bit rates ranged from 500 Mbps to 1.5 Gbps, PSNR 30-37 dB
  
• 
  10 bit to 8 bit conversion was done by rounding
  
• 
  True 8K sequences (not upsampled) would be desirable to judge without influence of the upsampling filter.
  

17.1.1.1.1.1.1.1.4JCTVC-C276 Screen content coding results using TMuC [C. Lan, J. Xu, F. Wu (Microsoft), G. Shi (Xidian Univ.)]

New coding tools had been presented for coding screen contents using an AVC framework in JCTVC-B084. This document presented new coding results using the TMuC, for intra coding and for random-access cases.

The following remarks and observations were recorded during the discussion:

• 
  Recorded at 10 fps
  
• 
  BD rate saving around 20% for random access inter, 28% for all intra.
  
• 
  More information about the characteristics of these sequences was requested – e.g., How often do frame changes occur, is there subpixel motion?
  
• 
  What would be typical test material? This may be the most difficult issue, as this has high influence on potential selection of tools.
  
• 
  As this is a new application area, the importance should be reviewed by the parent bodies.
  
• 
  Video (particularly news) includes lot of this type of content nowadays.