International organisation for standardisation organisation internationale de normalisation

4.3.2Contributions

4.3.2.1.1.1.1.1.1JCTVC-H0494 CE3: Chroma interpolation filter with half pixel phase offset [J. Lou, K. Minoo, L. Wang, A. Luthra (Motorola Mobility)]
4.3.2.1.1.1.1.1.2JCTVC-H0608 CE3: Cross-check for half-pel Chroma interpolation filter modification (JCTVC-H0494) by Samsung [E. Alshina, A. Alshin, J. H. Park (Samsung)] [late]
4.3.2.1.1.1.1.1.3JCTVC-H0182 CE3: Cross-check report for Motorola Mobility chroma interpolation filter (JCTVC-H0494) [T.Chujoh (Toshiba)]

4.4CE4: Quantization

4.4.1Summary

4.4.1.1.1.1.1.1.1JCTVC-H0034 CE4: Summary report of Core Experiment on Quantization [K Sato (Sony), M Budagavi (TI), M Coban (Qualcomm), H Aoki (NEC), X Li (MediaTek)]

The results of the CE are summarized as follows:

Subtest 1: Delta QP coding

Subtest 1.1 is the current WD5 (differential encoding relative to previous QP) wherever it came from.

Subtest 1.2 takes reference to left QP from same LCU (if available) otherwise last QP from same LCU – saves 0.1% for intra, 0.0% for inter.

Subtest 1.3 also only takes reference to QPs from same LCU, but makes it dependent on intra pred mode, or in inter case average of left & top depending on availability – saves 0.5% for intra, 0.3/0.2% for RA/LD.

Subtest 1.4 is weighted average depending on equality of neighboured QP values – saves approximately 0.2% on average .

Subtests 1.3 and 1.4 are too complex compared to the benefit they give.

Subtest 1.2 is slightly more complex than 1.1 and gives very little gain.

Conclusion: No change to WD.

Subtest 2: Quant matrix coding

There were 11 proposals, 6 are lossy or are configurable as lossy.

Subjective tests were asked to be done to identify whether lossy compression causes a problem.

A flat matrix was tested only if it is at the same bit rate.

It was requested to check whether the different proposals have the same bit rate, but they seemed likely to be in same range as tested.

Tests were done at QP 32 and 37.

It was suggested to select 3 lossy compression proposals and one lossless one.

A breakout group (coordinated by K. Sato and T. Suzuki) was asked to suggest a test setup and assess the complexity of the different lossy and lossless proposals (including memory size, e.g. using the same matrix values for groups of coefficients). The BoG reported its results in H0710.

Concern was expressed regarding whether the current set of matrices (6 in total) is sufficient.

After some subjective testing, it was planned to identify whether lossy compression causes problems.

After the H0710 BoG report was produced, subjective viewing was conducted as additional BoG activity, with results reported in H0730.

4.4.1.1.1.1.1.1.2JCTVC-H0710 BoG Report on Quantization Matrices [T. Suzuki, K. Sato]

This document reported the results of discussions on quantization matrices subjective viewing and complexity analysis.

As result of discussion of the CE4 report it had been agreed that subjective quality comparison should be conducted to determine the subjective difference between flat quantization matrices and lossless quantization matrix encoding, and the difference between lossless and lossy matrix encoding – and then after the viewing session, to assess the complexity of each proposal.

It was recommended that tests should be conducted immediately with the main purpose being to identify whether lossy QM coding is affecting the video quality. Planned viewing conditions were described.

It was agreed that if proposals originally planned for being subjectively viewed are not able to provide material for viewing by noon on 02-03, these proposals might be unable to be included in the test. It was remarked that the test should, at a minimum, be run with lossless and one lossy method (one of the more low-rate versions), and in addition should include a comparison against flat quantization matrix usage.

4.4.1.1.1.1.1.1.3JCTVC-H0730 BoG Report on Quantization Matrices Subjective Viewing [K. Sato, T. Suzuki, X. Zhang, R. Joshi, J. Zheng, M. Zhou]

Tested: Flat, Lossless, Lossy 2.1, 2.2, 2.4.

None of methods used compression for 4x4 and 8x8.

Two Sequences (BQ Mall, Kimono) were investigated at QP=32 only.

The lossless matrix encoding is usually judged better than flat, however typically it had a slightly higher bit rate.

There seemed to be no subjective difference between the lossless and the three lossy compression methods.

The three lossy methods have substantial differences in bit rates.

For "lossy" compression, data rate may be less of an argument than storage of matrices (e.g. using the matrix values for several coefficients by subsampling). Lossless compression of infrequently-used SPS information may also not be too relevant, whereas the storage of reconstructed values should be reduced.

If memory is of concern, should the default matrices also be downsampled?

Several experts also suggested to have non-uniform downsampling (with lower frequencies less downsampled).

It was also expressed that weighting is more important for larger transforms. No conclusion seemed possible from the current CE4 subtest 2 result.

A new design of CE4 should only target the "subsampling" (uniform and non-uniform) approaches for the matrices of 16x16 and 32x32 transforms – also to be investigated for default matrices.

No change was suggested for 4x4 and 8x8.

H0384, in the version where the 8x8 matrix is re-used as a sub-sampled matrix for the 16x16 and 32x32 cases, increases the data rate as compared to the lossless matrix case (approximately 10-15%), due to the fact that the quantization is not the same. However, H0384 is reported to have a second part where downsampled versions of the 16x16 and 32x32 original matrices were conveyed in the SPS header, which would not have visual impact and have very similar data rates (<5% difference, sometimes higher, sometimes lower). Linear interpolation is used for determining the final values. This has not been cross-checked in terms of subjective quality, but the cross-checker confirms that the BD performance is equal (0.1% difference). H0230 is similar but even simpler by using the same value for several neighbouring coefficients (i.e. zero order interpolation). It is reported that the difference in data rate compared to lossless QM encoding is very small except for BQ Mall RA (which would need some more clarification why this is the case).

Decision: Adopt the simple subsampling method (H0230 which uses separate 8x8 matrices with zero-order hold interpolation for 16x16 and 32x32, but keep DC as separate value which was not done in H0230).

This was further discussed in plenary session.

The allowed range of values for QM entries should be as in AVC (1 to 255).

Decision: Use the same subsampling for default matrices for 16x16 and 32x32, and make the defaults for 16x16 and 32x32 be the 8x8 defaults. Preliminary text, uploaded as a revision of H0230, was reviewed.

It was suggested for a CE or AHG to further investigate the following:

• 
  default values for each size
  
• 
  whether a denser sampling of the low frequency range is necessary
  

Note: Default matrix in hardware could be implemented differently than storing it in memory, but for software it would be desirable to implement it with the same subsampling as the loadable matrix.
Subtest 3: QP signalling on TU/PU basis instead of CU basis:

JCTVC-H0506/H0508 TU/PU level instead CU level (minimum 8x8).

Objective results: Loss in BD bit rate/PSNR, loss for TU based method (506) 0.5-1%, loss for PU based method (508) lower (0.1/0.2%).

Some results were shown with an assertion that it improves quality (still frames) – although this assertion seemed not fully obvious.

There was an argument that it is needed for rate control – argued that it is beneficial to have QP switching where motion vectors are switched (i.e. switching at the PU level would be needed here).

Instead of results from QP maps that are different from the current CE conditions (and with increased bit rate due to finer granularity), it would be better to compare, with the same maps, a CU-only versus a CU/PU based method (i.e. varying the smallest CU size).

It seemed unclear from the results whether the additional PU level signalling costs overhead or gives a benefit (compared to the same QP changes encoded at CU level only). The same granularity could be achieved by using 8x8 CUs with the current syntax, if needed. It would be anyway only for some specific applications.

No support was expressed except by proponents – and no action was taken.

JCTVC-H0260 reports PSNR gains by using different QP values for 8x8 and 4x4 TUs. As this is influencing the decision of RDO (actual bit rate varies by 2% typically).

It was asked whether this impacts visual quality. It was suggested that it may have some impact on rate control decisions, and may hurt perceptual quantization

No action taken.
Subtest 4 – Finer spatial granularity of QP selection.

The argument is that finer spatial granularity of QP selection is needed in HEVC because rate control may need to be aligned with CU boundaries which may be larger than the MBs in AVC. However, no rate control algorithm was currently available to give proven evidence about this.

The results indicate that it may be useful with constant QP but this is not relevant in practice.

No support for this was expressed by other experts in the absence of a real proof of value.