Joint Video Experts Team (jvet) of itu-t sg 6 wp and iso/iec jtc 1/sc 29/wg 11

CE2.3: SAO

Test#	Description	Document#
AhG13	Coding gain of SAO in VTM	JVET-K0013
2.3.1.a	SAO with EO sign constraints removal	JVET-K0233
2.3.1.b	SAO with EO sign constraints removal and more EO patterns	JVET-K0233
2.3.2.a	SAO Merge with 128×128 unit	JVET-K0324
2.3.2.b	SAO Merge with 128×128 or 256×256 unit	JVET-K0324
2.3.2.c	SAO Merge with 64×64 or 128×128 unit	JVET-K0324
2.3.3.a	SAO Palette results and discussion: Test a (fixed block size, as LCU)	JVET-K0192
2.3.3.b	SAO Palette results and discussion: Test b (adaptive block size)	JVET-K0192
2.3.4.a	SAO Modification: EO (modified threshold)	JVET-K0176
2.3.4.b	SAO Modification: BO offset quantization	JVET-K0176
2.3.4.c	SAO Modification: BO offset band (fixed instead of being signalled)	JVET-K0176
2.3.4.d	SAO Modification: shifted region	JVET-K0176
2.3.4.e	SAO Modification: EO and BO offset range	JVET-K0176
2.3.4.f	SAO Modification: All	JVET-K0176
2.3.4.g	SAO Modification: a + b + d	JVET-K0176
2.3.5	CTU adaptive sample adaptive offset (modified band position and range)	JVET-K0153

Results vs. VTM:

	AI					RA					LDB
Test#	Y	U	V	EncT	DecT	Y	U	V	EncT	DecT	Y	U	V	EncT	DecT
AhG13
2.3.1.a	0.00%	0.03%	0.03%	100%	100%	-0.04%	-0.10%	-0.07%	100%	101%	-0.17%	0.06%	-0.04%	100%	99%
2.3.1.b	-0.04%	-0.06%	-0.07%	101%	100%	-0.13%	-0.21%	-0.11%	100%	100%	-0.15%	-0.32%	-0.21%	100%	100%
2.3.2.a	-0.12%	-0.86%	-0.92%	99%	104%	-0.25%	-2.00%	-1.82%	101%	109%	-0.28%	-2.99%	-3.68%	101%	104%
2.3.2.b	-0.12%	-0.84%	-0.91%	99%	103%	-0.26%	-2.02%	-1.85%	100%	107%	-0.25%	-2.80%	-3.58%	100%	105%
2.3.2.c	-0.11%	-0.97%	-1.12%	93%	97%	-0.30%	-2.10%	-2.10%	100%	100%	-0.43%	-3.28%	-4.30%	96%	96%
2.3.3.a	-0.17%	-0.43%	-0.43%	100%	92%	-0.29%	-0.90%	-0.72%	100%	87%	-0.37%	-1.60%	-1.87%	96%	92%
2.3.3.b	-0.17%	-0.33%	-0.41%	102%	93%	-0.38%	-1.16%	-1.03%	97%	100%	-0.52%	-2.07%	-2.75%	98%	111%
2.3.4.a	-0.08%	-0.15%	-0.18%	99%	96%	-0.12%	0.08%	0.14%	97%	94%	0.03%	0.51%	0.66%	98%	95%
2.3.4.b	-0.04%	0.39%	0.51%	99%	97%	-0.04%	0.43%	0.51%	97%	97%	-0.05%	0.06%	0.38%	100%	103%
2.3.4.c	-0.06%	0.86%	1.17%	100%	98%	0.13%	0.97%	1.29%	98%	101%	0.49%	1.59%	2.82%	100%	101%
2.3.4.d	0.00%	-0.13%	-0.15%	99%	96%	-0.01%	-0.15%	-0.14%	97%	94%	0.02%	-0.20%	-0.49%	98%	95%
2.3.4.e	0.00%	0.00%	0.01%	100%	100%	-0.01%	0.00%	0.03%	98%	96%	0.02%	0.03%	-0.06%	101%	101%
2.3.4.f	-0.15%	0.63%	0.90%	100%	96%	-0.02%	1.06%	1.44%	97%	94%	0.46%	2.30%	3.71%	98%	95%
2.3.4.g	-0.13%	0.17%	0.23%	100%	98%	-0.17%	0.46%	0.62%	98%	98%	-0.06%	0.58%	1.14%	99%	97%
2.3.5	0.06%	-0.75%	-0.84%	100%	101%	-0.06%	-0.92%	-0.94%	100%	100%	-0.15%	-1.52%	-2.03%	100%	101%

Results vs. BMS:

	AI					RA					LDB
Test#	Y	U	V	EncT	DecT	Y	U	V	EncT	DecT	Y	U	V	EncT	DecT
AhG13
2.3.1.a	0.00%	-0.01%	-0.07%	100%	100%	0.02%	-0.08%	-0.06%	100%	100%	-0.07%	-0.07%	-0.04%	100%	100%
2.3.1.b	-0.03%	-0.03%	-0.12%	100%	100%	-0.03%	-0.24%	-0.17%	100%	100%	-0.14%	-0.57%	-0.30%	101%	100%
2.3.2.a	-0.09%	-0.74%	-0.69%	99%	102%	-0.15%	-2.34%	-2.19%	101%	105%	-0.37%	-3.37%	-3.99%	100%	103%
2.3.2.b	-0.10%	-0.70%	-0.67%	98%	101%	-0.17%	-2.36%	-2.20%	99%	104%	-0.42%	-3.53%	-4.27%	99%	103%
2.3.2.c	-0.08%	-0.87%	-0.92%	98%	98%	-0.18%	-2.43%	-2.40%	101%	103%	-0.42%	-3.77%	-4.60%	99%	101%
2.3.3.a	-0.15%	-0.13%	0.00%	99%	94%	-0.18%	-1.10%	-0.87%	101%	92%	-0.47%	-1.99%	-2.43%	101%	97%
2.3.3.b	-0.14%	-0.02%	0.07%	101%	96%	-0.24%	-1.38%	-1.23%	100%	95%	-0.56%	-2.60%	-3.15%	105%	97%
2.3.4.a	0.00%	0.11%	0.02%	100%	98%	0.01%	0.10%	0.05%	98%	98%	-0.05%	0.21%	0.71%	101%	97%
2.3.4.b	-0.04%	0.63%	0.70%	101%	105%	0.02%	0.50%	0.59%	100%	101%	-0.11%	0.03%	0.30%	101%	103%
2.3.4.c	-0.07%	1.22%	1.43%	102%	106%	0.08%	0.83%	1.05%	100%	102%	0.09%	1.26%	2.18%	101%	105%
2.3.4.d	0.00%	-0.13%	-0.13%	100%	98%	0.01%	-0.17%	-0.13%	98%	98%	-0.07%	-0.41%	-0.19%	99%	97%
2.3.4.e	0.00%	0.00%	0.00%	100%	100%	0.02%	-0.01%	0.01%	99%	99%	-0.09%	0.01%	0.07%	100%	100%
2.3.4.f	-0.08%	1.37%	1.49%	101%	100%	0.04%	0.99%	1.14%	97%	99%	0.09%	1.80%	3.39%	98%	97%
2.3.4.g						-0.01%	0.59%	0.67%	98%	99%
2.3.5	0.07%	-0.90%	-1.02%	100%	102%	0.02%	-0.88%	-0.96%	100%	100%	-0.13%	-1.62%	-1.99%	100%	101%

2.3.2.x and 2.3.3.x have picture-level optimization and give gain up to 0.4% in RA compared to VTM, and It was remarked that the same effect could be achieved by a smarter encoder e.g. performing lookahead

The methods that do not use picture level optimization provide small gain (<0.2% for RA compared to VTM)

Compared to BMS, the methods without picture level opt. do not provide gain any more in RA, and the gain of the picture-level methods reduces to <0.25%

Generally, the gain in LDB is slightly higher (around 0.5% vs VTM and BMS for picture level opt.), where however this would introduce one frame delay at the encoder.

The results do not suggest any action at this moment. a) No evidence is currently available that there is subjective improvement; b) when combined with ALF, most gain is lost; c) more smart encoders likely could get similar gain by picture-level optimization (at the expense of additional encoding delay) without syntax change.