International organisation for standardisation

5.8.2Contributions

5.8.3Subtest 1: Block level filter adaptation with directional feature (BADIR)

5.8.3.1.1.1.1.1.1JCTVC-F301 CE8 Subtest 1: Block-based filter adaptation with features on subset of pixels [P. Lai, F. C. A. Fernandes, E. Alshina, I.-K Kim (Samsung)]
5.8.3.1.1.1.1.1.2JCTVC-F223 CE8.1: Verification results of Samsung's Proposal (JCTVC-F301) [Faouzi Kossentini, Hsan Guermazi (eBrisk)]
5.8.3.1.1.1.1.1.3JCTVC-F387 CE8.1: Cross-check result of Samsung's adaptive loop filter (JCTVC-F301) [T. Ikai, Y. Yasugi (Sharp)]

5.8.3.1.1.1.1.1.4JCTVC-F321 CE8.1: Block based Adaptive Loop Filter by MediaTek, Qualcomm and Toshiba [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)]

5.8.3.1.1.1.1.1.5JCTVC-F307 CE8 Subtest 1: Cross-check of MediaTek, Qualcomm and Toshiba’s proposal (JCTVC-F321) on block based ALF and temporal prediction of ALF coefficients [P. Lai, F. C. A. Fernandes (Samsung)]
5.8.3.1.1.1.1.1.6JCTVC-F516 CE8 Subtest 1: Cross verification of MQT's proposal by Intel [Y. Chiu, L. Xu, W. Zhang, Y. Han (Intel)]
5.8.3.1.1.1.1.1.7JCTVC-F384 CE8.1:Block based Adaptive Loop Filter with flexible syntax and additional BA mode by Sharp and Qualcomm [T. Ikai (Sharp), M. Karczewicz, I. S. Chong (Qualcomm), Y. Yasugi, T. Yamazaki (Sharp)]
5.8.3.1.1.1.1.1.8JCTVC-F067 CE8 Subtest 1: Crosscheck for Sharp and Qualcomm's Adaptive Loop Filter in JCTVC-F384 [C.-Y. Chen, Y.-W. Huang (MediaTek)]
5.8.3.1.1.1.1.1.9General

Summary table results against HM3.0 16-pass anchor (mandatory) below

	JCTVC-F301 Samsung				JCTVC-F321 Toshiba/Qualcomm/MediaTek				JCTVC-F384 Sharp/Qualcomm
	AI	RA	LB	LP	AI	RA	LB	LP	AI	RA	LB	LP
BD-rate	0.1	0.2	0.2		0.0	−0.3	−0.3	−0.2	−0.3	−0.3	−0.4	−0.4
Enc. Time	99	100	100		100	101	100	101	101	100	100	101
Dec. time	96	98	98		100	105	101	102	102	100	101	102
Vertical taps	7				7				7

Table for combination of JCTVC-F321 and JCTVC-F384 (from JCTVC-F522, outside CE contribution):

	HM3.0 16-pass				HM3.1-dev-adcs
	AI	RA	LB	LP	AI	RA	LB	LP
BD-rate	−0.3	−0.7	−0.6	−0.7	−0.2	−0.5	−0.6	−0.6
Enc. Time	100	99	101	103	101	101	101	99
Dec. time	98	108	104	104	99	102	102	103
Vertical taps	7				5

The gains of these techniques therefore seem basically additive.

JCTVC-F384 reportedly introduces an extra mode for the encoder to choose from, thus increasing search complexity (in the absence of suitable fast decision-making techniques).

It was commented that it may be possible to improve upon the JCTVC-F301 design. However, such a proposal to improve it could be submitted in the future regardless of whether we take action now.

The text and software impact of JCTVC-F301 seems relatively minor.

Decision: Adopt JCTVC-F301 (a simplification of filter selection).

Regarding JCTVC-F321, it enables prediction of the coefficients from a previous frame based on a coded reference index. Actually, JCTVC-F321 has some other aspects as well, but the gain is mostly from this.

A second aspect of JCTVC-F321 involves having an increased number of pixel classification classes. The cross-checker reported the results separately in the spreadsheet results for JCTVC-F307. That aspect had little benefit, and increases complexity. That aspect was not adopted.

It must not violate the layering structure in the use of the reference index.

The proposal seems to enable more frequent use of ALF, as it reduces the per-picture cost of enabling it. Note that ALF has subjective benefits that likely exceed its objective benefit.

However, the benefit is rather small.

It was later suggested that the ability to choose the temporal prediction may somewhat help with encoder complexity (although this was not tested in the CE).

It was noted that using the PPS to send the coefficients already enables sharing of coefficients across frames.

Another aspect of JCTVC-F384 proposes a 2D class merging scheme rather than a 1D scan order based merging scheme. However, the benefit is very small, and the ordering in the edge direction dimension seems arbitrary.

No action was taken on JCTVC-F321 and JCTVC-F384.

It was noted that a slice parameter set scheme (JCTVC-E281 / JCTVC-F187) could also enable another type of sharing.

A participant suggested that rather than having a PPS index in the slice parameter set, there could be two indexes in the slice header – a PPS index and a slice parameter set index – each coded as ue(v).

Decision: Agreed – syntax and semantics was drafted in side activity for approval.

Can the APS index vary within a picture? Not for now.

See discussion of JCTVC-F747 for further notes relating to the APS.