International organisation for standardisation

5.8.7Subtest 5: Chroma filter size and shape

5.8.7.1.1.1.1.1.1JCTVC-F042 CE8 Subtest 5: Adaptive Loop Filtering of Luminance and Chrominance Samples Using Same Filtering Shape, Structure and Map [F. Kossentini, H. Guermazi, N. Mahdi, M. A. Ben Ayed, M. Horowitz (eBrisk)]

Luma and chroma filter shapes are the same as those of HM3.1-dev-adcs, but with smaller extent. When CU adaptive luma filtering is applied, chroma filtering follows the luma filtering control.

5.8.7.1.1.1.1.1.2JCTVC-F188 CE8.5: Verification results of eBrisk Video's Proposal JCTVC-F042 [T. Yamakage, T. Watanabe (Toshiba)]

Source code was studied as well as compiled and run.

5.8.7.1.1.1.1.1.3JCTVC-F157 CE8.5: Unified Chroma Filter Shapes with Luma Shapes for ALF [T. Yamakage, T. Watanabe, T. Chujoh (Toshiba), C.-Y. Chen, C.-M. Fu, C.-Y. Tsai, Y.-W. Huang, S. Lei (MediaTek), M. Karczewicz, I. S. Chong (Qualcomm)]

The modification point from the HM3.0 software (5x5 rectangle shape chroma filter) is to use of a 7x7 diamond shape or RDO from 5x5, 7x7 and 9x7 diamond shapes.

5.8.7.1.1.1.1.1.4JCTVC-F245 CE8 subtest 5: MQT chroma ALF (JCTVC-F157) cross-check [M. Budagavi (TI)] [late upload 07-10]

Source code was reviewed and confirmed that it was consistent with the description in JCTVC-F235.

5.8.7.1.1.1.1.1.5JCTVC-F454 CE8 subtest 5: Cross-check results for proposal JCTVC-F157 [S. Esenlik, M. Narroschke (Panasonic)]

Panasonic cross-checked 2 of the 3 variants of the proposal which are implemented using HM3.0 as the base. The recognized differences are as follows:

• 
  Variant 1: 5x5 square filter shape of chroma ALF in HM3.0 is replaced by 7x7 diamond filter shape.
  
• 
  Variant 2: 5x5 square filter shape of chroma ALF in HM3.0 is replaced by 3 diamond shaped filters (5x5, 7x7 and 9x9) and a rate-distortion optimization method is implemented for filter size decision.
  

5.8.7.1.1.1.1.1.6JCTVC-F235 CE8, Subset 5, Tool 3: Chroma ALF with reduced vertical filter size (JCTVC-E287) [M. Budagavi, V. Sze, M. Zhou (TI)]

The contribution presents two ALF filters that reduce vertical size of chroma ALF filter. The following shows these filters, where all two filters have a vertical size of 3 instead of 5 in HM3.0.

5.8.7.1.1.1.1.1.7JCTVC-F666 CE8.5: Verification results of TI's Proposal [Faouzi Kossentini] [late reg. 07-06, upload 07-09]

5.8.7.1.1.1.1.1.8General

Table of results against HM3.0 16-pass High Efficiency anchor (mandatory)

	JCTVC-F042 eBrisk				JCTVC-F157 Toshiba/MediaTek/Qualcomm 7x7 diamond				JCTVC-F235 TI 9x3 rectangle
	AI	RA	LDB	LDP	AI	RA	LDB	LDP	AI	RA	LDB	LDP
BD-rate Y	0.2	0.2	0.3	0.3	0.0	0.0	0.0	0.0	0.0	0.0	0.0
BD-rate U	−0.1	−0.5	−0.9	−1.1	−0.1	−0.4	−0.4	−0.4	0.1	0.3	0.1
BD-rate V	−0.1	−0.3	−0.8	−1.3	−0.1	−0.4	−0.5	−0.8	0.2	0.3	−0.2
Enc. Time	102	101	101	101	100	100	100	100
Dec. time	91	95	95	93	99	99	100	100	104	108	102
Vertical taps	5				7				3

continued:

					JCTVC-F157 Toshiba/MediaTek/Qualcomm 5x5, 7x7, 9x7 RDO				JCTVC-F235 TI 9x3 diamond
	AI	RA	LDB	LDP	AI	RA	LDB	LDP	AI	RA	LDB	LDP
BD-rate Y					0.0	0.0	0.0	0.0	0.0	0.0	0.0
BD-rate U					−0.5	−0.9	−0.4	−0.7	0.1	0.3	0.2
BD-rate V					−0.6	−0.8	−0.7	−1.0	0.3	0.3	0.1
Enc. Time					100	100	100	100	104	108	102
Dec. time					101	100	101	100	94	91	95
Vertical taps					7				3

Decision: Use the same shapes as for luma (proposed in both JCTVC-F303, making the shapes consistent as recommended in JCTVC-F042 and JCTVC-F157). For chroma, send:

• 
  an enable/disable flag
  
• 
  if enabled, a flag that selects the shape
  

Further study may be desirable for shrinking the chroma region of support for 4:2:0 and 4:2:2.

Regarding the JCTVC-F042 proposal to control the chroma filtering together with the luma filtering for CU based decision-making – for further study.

It was remarked that there is some relationship between gains for AIF and ALF, and that the complexity of each of them may interact with the behaviour of the other.

5.8.8Discussion and Conclusions

See above.

5.9CE9: MV coding and skip/merge operation (37)

5.9.1Summary

5.9.1.1.1.1.1.1.1JCTVC-F029 CE9: Summary report of Core Experiment on MV coding and skip/merge operations [Y.-W. Huang, B. Bross, M. Zhou, W.-J. Chien, I.-K. Kim (CE coordinators)] [upload 07-13]

Some descriptions of CE plans were not fully precise, and some aspects were modified after the original CE plan was established.

Category 1: Unification of AMVP and Merge candidate lists

Three methods in that area were planned to be tested, and one was withdrawn, so UNI01 and UNI03 are the remaining candidates.

It was suggested to focus on the UNI03 method in this category.

JCTVC-F297 describes UNI03. The CE description of this scheme was somewhat vague, with details determined by a proponent in the interim – although the participants were satisfied with the result.

This was proposed as a harmonization / clean-up of the design with some complexity reduction and some minor gain (roughly 0.2%).

Category 2: Simplification of MVP list construction

It was suggested to select the SP01, SP04, SP06S2, SP07 simplification proposals.

Category 3: Improvements of coding efficiency

No improvements verified.

Category 4: Evaluation of inferred merge (a.k.a. "partial merge")

Relates to part of JCTVC-F465 (and JCTVC-F082).

It was remarked that the "PART14" scheme was not described in the CE plan.

Decision: Remove inferred merge

Decision: Add SPS flag to remove 4x4 inter prediction, and set the flag to disable 4x4 for common conditions (B. Bross responsible for delivery of the software change).

Remark: Will there be a perceptual effect from removing this?

Further study was suggested to determine whether that should be used in the future.

Category 5: Parsing robustness for AMVP and merge

Non-CE contributions JCTVC-F068 JCTVC-F341 JCTVC-F347 JCTVC-F402 JCTVC-F470 JCTVC-F474 were noted to be closely related.

ROB04 has two elements

• 
  Replace redundant MVPs with some predefined values
  
• 
  Something else is done to remove some redundant MVPs without using their values to identify them.
  

It was remarked that this scheme has some complexity issues

ROB02 is asserted to be simpler, although it has more coding loss.

ROB06, a combination of ROB04 and ROB02, has less loss, but was not a CE proposal.

A BoG coordinated by B. Bross was set up to review the parsing fixes and various other above-described aspects. See BoG report JCTVC-F744.