International organisation for standardisation

5.8.6Subtest 4: Filter size and shape

5.8.6.1.1.1.1.1.1JCTVC-F041 CE8 Subtest 4: Adaptive Loop Filtering Using Two Filter Shapes [F. Kossentini, H. Guermazi, N. Mahdi, M. A. Ben Ayed, M. Horowitz (eBrisk)]

ALF employs two filter shapes, diamond 7×5 with 12 coefficients (TI proposal) and Cross 11x5 with 8 coefficients (eBrisk proposal). This reduce the worst-case number of coefficients from current 20 (diamond 9x7) to 12. Also both are vertical size 5, thus reducing memory bandwidth.

Same as in HM3.1-dev-adcs-adcs anchor, encoder determines filter shape first (using block-based classification), then determining the BA or RA, then CU-on/off control map.


5.8.6.1.1.1.1.1.2JCTVC-F305 CE8 Subtest 4: Cross-check of eBrisk proposal (JCTVC-F041) on adaptive loop filtering using two filter shapes [P. Lai, F. C. A. Fernandes (Samsung)]
5.8.6.1.1.1.1.1.3JCTVC-F234 CE8, Subset 4, Tool 3: ALF decode with reduced vertical filter size (JCTVC-E060) [M. Budagavi, V. Sze, M. Zhou (TI)]

Nx5-Set1 ALF filter set (5x5, 7x5, or 9x5) that reduces the vertical size of the filter to 5. Since this filter operate on only five lines instead of the original seven, the size of the line buffers goes down from 6 lines to 4 lines. This was compared with the HM3.0 anchor which adopts 5x5, 7x7, or 9x7 filter.

5.8.6.1.1.1.1.1.4JCTVC-F304 CE8 Subtest 4: Cross-check of TI’s proposal (JCTVC-F234) on ALF decode with reduced vertical filter size [P. Lai, F. C. A. Fernandes (Samsung)]

5.8.6.1.1.1.1.1.5JCTVC-F303 CE8 Subtest 4: ALF using vertical-size 5 filters with up to 9 coefficients [P. Lai, F. C. A. Fernandes (Samsung), H. Guermazi, F. Kossentini (eBrisk)]

Two filter shapes, star-5x5 with 9 coefficients and cross-11x5 with 8 coefficients are used for luma ALF. This reduce the worst-case number of coefficients from current 20 (diamond 9x7) to 9. Also both are vertical size 5, thus reducing memory bandwidth.

Same as in HM3.1-dev-adcs-adcs anchor, encoder determines filter shape first (using block-based classification), then determining the BA or RA, then CU-on/off control map.

5.8.6.1.1.1.1.1.6JCTVC-F324 CE8 Subset 4: Cross-Verification of Samsung and eBrisk adaptive loop filter (JCTVC-F303) by Qualcomm [I. S. Chong, M. Karczewicz (Qualcomm)]

5.8.6.1.1.1.1.1.7JCTVC-F244 CE8: Samsung ALF (JCTVC-F303) crosscheck [M. Budagavi (TI)] [late upload 07-10]
5.8.6.1.1.1.1.1.8General

High-efficiency test results:

	JCTVC-F041 eBrisk Anchor: HM3.1-dev-adcs {dia-7x5, cross-19x5}				JCTVC-F234 TI Anchor: HM3.0 {9x7, 7x7, 5x5}				JCTVC-F303 Samsung/eBrisk Anchor: HM3.1-dev-adcs {dia-7x5, cross-19x5}
	AI	RA	LDB	LDP	AI	RA	LDB	LDP	AI	RA	LDB	LDP
BD-rate	0.0	0.0	0.0	0.0	0.1	0.2	0.3		0.1	0.2	0.1
Enc. Time	96%	97%	97%	97%					82%	92%	93%
Dec. time	100%	100%	99%	99%	100%	93%	99%		98%	98%	97%
Vertical taps	5				5				5
Max# of coeff.	12				17				9

Note that the anchor is different for different proposals in the table.

Each has a vertical extent of 5. The shapes are otherwise somewhat different.

JCTVC-F303 has the lowest complexity among these, from both the encoder and decoder perspective, and little penalty.

Decision: Adopt JCTVC-F303.