International organisation for standardisation organisation internationale de normalisation

4.2CE2: Chroma RQT depth

4.2.1Summary

4.2.1.1.1.1.1.1.1JCTVC-H0032 CE2: Summary report of Core Experiment on Chroma RQT depth [K. Sugimoto, X. Guo, X. Wang (CE Coordinators)]

The following tools were tested:

test	Proposal Document	Proponent	Tool description	Cross-checker & Document #
4.1 JCTVC-G283	JCTVC-H0317	MediaTek PKU	- Signalling Max chroma RQT depth for intra - Chroma TU split for inter same as done in luma with max depth=1 - no 32x32 transform for chroma	Qualcomm JCTVC-H0543
4.2 JCTVC-G442	JCTVC-H0235	Mitsubishi NHK	- Implicit chroma RQT depth depending on intra and inter PU mode - allow 32x32 transform for chroma	MediaTek JCTVC- H0319
4.3 JCTVC-G980	JCTVC- H0538	Qualcomm	- Signalling Max chroma RQT depth for intra - Chroma TU split for inter same as done in luma with max depth=1 - allow 32x32 transform for chroma	Mitsubishi JCTVC- H0236
4.4.0	JCTVC- H0308	Harmonization	- Signalling Max chroma RQT depth for intra and inter - no 32x32 transform for chroma	HiSilicon JCTVC- H0258
4.4.1	JCTVC- H0308	Harmonization	- Signalling Max chroma RQT depth for intra and inter - allow 32x32 transform for chroma	Panasonic JCTVC-H0122

The harmonized solution H0308, in principle, includes elements from all 3 original proposals (G283, G442, G980), and it was decided by reflector discussion to also test G442 separately. Therefore, only H0308 and H0235 (the previous G442) needed to be considered. Among the CE participants, it was suggested to adopt H0308 without the 32x32 chroma transform (configuration 4.4.0)

The approach is to add signalling (2 flags for intra and inter) at the SPS level for limitation of chroma RQT depth (a similar approach as currently for luma and chroma combined). The maximum RCT depth for chroma would still be the same as currently, and therefore this is not a decoder complexity reduction. What was tested in 4.4.0 was using intra chroma RQT depth = 1 and inter chroma RQT depth = 2 (not sequence adaptive), and luma RQT depth = 3 as in common test conditions. With fixed restricted chroma RQT depth (not sequence adaptive), slight gain on chroma (1–2%) is reported.

One comment: An encoder could not decide how to make it adaptive per sequence when it is signalled at the SPS level.

Several experts expressed concerns over whether it is necessary to signal the chroma depth depth at all, suggesting to make it fixed relative to luma. Restriction of chroma RQT depth (without signalling) seems to be useful and if made non-adaptive could even mean a decoder complexity reduction. It appears that either luma depth minus one (as from H0317) or minus two may make sense, but the latter case was not studied. Further it is not known how the performance would be with luma RQT depth other than 3. CE participants will investigate this – also in context of consideration of H0123.

4.2.1.1.1.1.1.1.2JCTVC-H0731 CE2: Additional tests on reducing Max chroma RQT depth [X. Zhao, J. An, X. Guo, S. Lei (MediaTek), L. Guo, X. Wang, M. Karczewicz (Qualcomm), A. Minezawa, K. Sugimoto, S. Sekiguchi (Mitsubishi), A. Ichigaya, S. Sakaida (NHK), S. Ma, W. Gao (PKU)] [late 02-06]

An aspect raised in the discussion was about the effect on spatial prediction, As the size of spatial prediction is derived from transform size, this may become difficult in cases where the chroma prediction is using luma, i.e. the LM mode.

From that point of view, worst-case complexity reduction is not an argument, and in principle the only advantage is compression gain.

Combination with H0123 was implemented, with simulations still running when discused.

Track B initially recommended to adopt the result from CE2 such that (without signalling in SPS) the version from H0731 where chroma depth is luma depth -1 and luma depth -2 in inter cases (minimum 1). This was then made subject to verification of combination with H0123. The topic was then further discussed in JCT-VC plenary.

Note: It was confirmed that max chroma TU size should be kept as 16x16. This restriction is intended only for the 4:2:0 case and may be revised in the future for 4:2:2, 4:4:4, etc.

In later discussion, the potential impact on non-4:2:0 chroma formats was discussed. There was also some concern expressed about the potential cases where a region covered by multiple luma blocks would be required to be coded with a single chroma block. It was noted that there had not been subjective quality evaluation of this, and that we should be past the stage of adopting changes for purposes of very small coding efficiency improvement.

After initially being recommended for adoption in Track B, it was decided not to adopt this change.
Average BD bit rate gain for each test is shown below.

test	High Efficiency						Low Complexity						High 10
	Intra		Random Access		Low Delay		Intra		Random Access		Low Delay		Random Access
	Y	UV	Y	UV	Y	UV	Y	UV	Y	UV	Y	UV	Y	UV
4.1	0.02	-2.47	-0.05	-1.32	-0.15	-0.02	0.02	-1.92	-0.01	-1.39	-0.02	-0.93	0.02	-2.59
4.2	-0.05	-2.46	-0.06	-1.43	-0.07	-1.14	0.02	-2.24	-0.01	-1.59	-0.03	-2.16	-0.02	-4.45
4.3	-0.05	-2.46	-0.09	-1.35	-0.10	-0.53	0.02	-2.24	-0.01	-1.72	-0.05	-2.13	-0.03	-4.08
4.4.0	0.02	-2.47	-0.04	-1.61	-0.02	-0.92	0.02	-1.92	0.00	-1.40	-0.01	-1.11	-0.01	-2.93
4.4.1	-0.05	-2.46	-0.05	-1.67	-0.04	-1.21	0.02	-2.24	0.00	-1.61	-0.02	-1.38	-0.01	-4.29

Average runtime for encoding and decoding compared with the anchor are shown below.

test	High Efficiency						Low Complexity						High 10
	Intra		Random Access		Low Delay		Intra		Random Access		Low Delay		Random Access
	ENC	DEC	ENC	DEC	ENC	DEC	ENC	DEC	ENC	DEC	ENC	DEC	ENC	DEC
4.1	100%	99%	97%	100%	98%	100%	101%	99%	98%	100%	99%	100%	98%	100%
4.2	99%	99%	96%	100%	97%	100%	100%	99%	97%	100%	98%	100%	96%	100%
4.3	99%	100%	97%	100%	97%	100%	98%	100%	98%	100%	98%	100%	97%	99%
4.4.0	99%	100%	98%	100%	98%	100%	99%	100%	98%	100%	99%	100%	97%	99%
4.4.1	98%	100%	99%	100%	99%	99%	98%	100%	99%	100%	99%	99%	98%	100%

4.2.2Contributions

4.2.2.1.1.1.1.1.1Test 1

4.2.2.1.1.1.1.1.2JCTVC-H0317 On Chroma Cbf Signalling [L. Zhao, X. Guo, S. Lei (MediaTek), S. Ma, W. Gao (PKU), D. Zhao (HIT)] [late]

4.2.2.1.1.1.1.1.3JCTVC-H0543 CE2: Cross verification of subtest 4.1 [L. Guo (Qualcomm)] [late]
4.2.2.1.1.1.1.1.4Test 2

4.2.2.1.1.1.1.1.5JCTVC-H0235 CE2 4.2: report on fixed TU depth for chroma [K. Sugimoto, A. Minezawa, S. Sekiguchi (Mitsubishi), A. Ichigaya, S. Sakaida (NHK)]

4.2.2.1.1.1.1.1.6JCTVC-H0319 CE2 subtest 4.2: Crosscheck Report [X. Zhao, X. Guo (MediaTek)] [late]
4.2.2.1.1.1.1.1.7Test 3

4.2.2.1.1.1.1.1.8JCTVC-H0538 CE2 4.3: Report on Limiting Chroma Transform Depth in RQT [L. Guo, X. Wang, M. Karczewicz (Qualcomm)]

4.2.2.1.1.1.1.1.9JCTVC-H0236 CE2 4.3: cross-verification of signalling max chroma TU depth [K. Sugimoto, A. Minezawa, S. Sekiguchi (Mitsubishi)] [late]
4.2.2.1.1.1.1.1.10Test 4

4.2.2.1.1.1.1.1.11JCTVC-H0308 CE2: Subset 4.4: Test Results on Harmonization of JCTVC-G283, JCTVC-G442 and JCTVC-G980 [X. Zhao, J. An, X. Guo, S. Lei (MediaTek), L. Guo, X. Wang, M. Karczewicz (Qualcomm), A. Minezawa, K. Sugimoto, S. Sekiguchi (Mitsubishi), A. Ichigaya, S. Sakaida (NHK), S. Ma, W. Gao (PKU)]

4.2.2.1.1.1.1.1.12JCTVC-H0122 CE2: Cross Check of Test 4.4.1 [Y. Shibahara, T. Nishi (Panasonic)] [late]
4.2.2.1.1.1.1.1.13JCTVC-H0258 CE2: Cross-verification of CE2 subtest4 [X. Zheng (HiSilicon)] [late]