Of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

SCE5.2
Motion data compression and memory reduction
5.2.1	JCTVC-M0283 (Sony, I2R) “SCE5 5.2.1 on Effectiveness of temporal motion vector prediction (TMVP)”	JCTVC-M0128 (JVC Kenwood)
5.2.2	JCTVC-M0141 (Sony, LG) “SCE5: Result of SCE5.2.2”	JCTVC-M0392 (MediaTek)
	JCTVC-M0142 (Sony) “SCE5: Buffer Size Estimation of SCE5.2.x”

The tests results for both categories are summarised in the next table.

Test	Short description	RA-2x	RA-1.5x	RA SNR	LP-2x	LP-1.5x	LP SNR	LB-2x	LB-1.5x	LB SNR	Enc	Dec
5.1.1 (Qualcomm)	insert BL MV into the completed list	0.25	0.19	0.29	0.26	0.21	0.32				100%	100%
5.1.2 (Qualcomm)	Inferred mode based on 8x8 sub-block	0.03	0.00	0.03	0.01	0.01	0.02				101%	103%
5.1.3		Withdrawn
5.1.4 (KDDI)	1. BL MV as a 3rd candidate	−0.03	0.17		−0.22	−0.06					100%	100%
	2. MV refinement	−0.05	0.30		−0.23	−0.02					100%	106%
	3. Test 1 + uncompressed MV	−0.84	−0.19		−0.89	−0.35					100%	100%
5.1.5 (MediaTek)	Simplified pruning	0.03	0.01	0.02	0.01	0.03	0.03				100%	100%
	2 BL MVs with simplified pruning	−0.61	−0.23	−0.61	−0.40	−0.16	−0.49				100%	100%
5.1.6 (Nokia)	1. use TMVP scaling	−0.03	0.00	0.00	−0.03	0.00	0.00				-	-
	2. w/o BL MV pruning	0.30	0.21	0.32	0.29	0.23	0.33				-	-
	3. Test 1 + Test 2	0.27	0.21	0.32	0.27	0.23	0.33				-	-
5.1.7(Intel)	Inferred mode based on NxN	0.09	0.08	0.12	0.09	0.08	0.11	0.09	0.06	0.10	103%	98%
5.1.8 (LG)	1. BL MV for AMVP	−0.04	−0.10	−0.11	−0.03	−0.04	−0.05				100%	100%
	2. Test 1 + uncompressed MV	−1.26	−0.64	−0.63	−1.07	−0.47	−0.69				100%	100%
5.1.9		Withdrawn
5.1.10 (ETRI)	1. Simplified BL MV pruning	0.08	0.05	0.07	0.09	0.07	0.08				100%	100%
	2. BL MV for AMVP	0.02	−0.10	−0.12	0.09	0.01	−0.02				100%	100%
5.2.1 (I2R, Sony)	1. disabling TMVP	0.76	0.34	0.38	0.92	0.57	0.48	0.81	0.46	0.44	100%	98%
	2. Test 1 + uncompressed MV	−0.49	−0.20	−0.17	−0.27	0.06	−0.25				-	-
	3. Test 1 + 8x8 compressed MV	−0.15	−0.12	−0.11	0.13	0.15	−0.14				101%	100%
5.2.2 (Sony, LG)	1. uncompressed MV	−1.10	−0.50	−0.50	−0.94	−0.41	−0.61				99%	99%
	2. 8x8 compressed MV	−0.79	−0.42	−0.45	−0.61	−0.34	−0.52				100%	101%

5.1.6.1 replaces the division currently used in BL MV scaling by the scaling operation of TMVP.

This is in principle supported by several experts, but it is also expressed that there may be a problem with the suggested solution when arbitrary scaling would be used. JCTVC-M0133 was also related to this topic and claimed to solve the issue. See notes on M0133 and BoG report M0449.

Note With 1.5X and 2X, it could not be expected that the results of division or TMVP scaling would be different; the 0.03% difference reported above is likely due to a bug which was fixed in SHM1.1

Note: Using uncompressed MV is known to give approx. 0.7% BR reduction.

5.1.2 and 5.1.7 (mode inferring from BL) do not provide benefit

Current anchor needs 5 additional comparisons (in worst case when all spatial candidates are available) in pruning for merge, compared to SL HEVC.

5.1.1 and 5.1.6 case 2 are avoiding additional pruning, but could end up with duplicate candidates (resulting in 0.2-0.3% BR increase)

For test 1: Prepare a table for each proposal, analyzing the number of additional or reduced operations (number of comparisons or if applicable, any other), compared to SHM.

An update of the report has been provided where the table above was augmented by the number of comparison operations. A graph is also included which plots the reduction in comparisons versus compression benefit.

a) 5.1.4.3 and 5.1.8.2 are using uncompressed BL motion (not to be considered)

b) 5.1.8.1 and 5.1.10.2 increase number of operations without significant benefit

c) 5.1.7 is same number of operations but produces loss

d) 5.1.5.2 reduces number of comparisons in pruning, but further deviates from the conventional pruning process in using a second base layer MV, which requires one more MV memory access and scaling operation. It provides 0.4% BR reduction, but M0112 from non-CE category seems to be a more attractive solution for further investigation, therefore it might be premature to adopt 5.1.5.2.

e) 5.1.5.1 (reduce 3 op, no loss), 5.1.10.1 (reduce 4 op, 0.1% loss) and 5.1.1/5.1.6 (reduce 5 op, 0.25% loss)

f) 5.1.4 has not reported on SNR scalability, but is expected to get loss in this case

Decision: Adopt 5.1.5.1 (simplified pruning for use of MV of BL as candidates in EL figure 2b from JCTVC-M0124) as modification of TextureRL scheme in SHM (not WD).
Test 2.2: Using 8x8 compressed BL MV (instead of 16x16 as in current SHM) gives around 0.5% bit rate reduction; 4x4 would give about 0.7%. JCTVC-M0142 provides analysis about additional memory.

Test 2.1: Disabling TMVP in the enhancement layer only has a loss of 0.7-0.8% BR reduction, whereas in a single layer configuration it should be 2% or more; disabling TMVP in enhancement layer and re-using the memory for better representation of the base layer MV (5.2.1 test 2) could lead to no performance loss, whereas saving MV memory for EL. More numbers about actual memory increase/saving were provided in a v4 of JCTVC-M0142 (see also additional notes under JCTVC-M0142).

For the case of 4K pictures, the additional storage necessary for 8x8 base layer MV would be around 1 MByte. The same amount of storage would be saved in case when TMVP is disabled in EL, and 8x8 base layer MV is used.

The best tradeoff seems to be usage of 8x8 compression, where we can

• 
  either increase memory by 1 Mbyte, and save 0.5% rate on average
  
• 
  or reduce memory by 1 Mbyte, without relevant loss.
  

The second bullet seems to indicate that the current usage of MV memory is not optimum in scalable enhancement layer.

One expert points out that by disabling TMVP in case of cropped base layer (i.e. some enhancement layer area without base layer reference), loss might occur.

A conclusion may be difficult to achieve at this meeting, but further study on the topic is necessary. Also the relation with refidx should be considered.

JCTVC-M0122, JCTVC-M0144 are also related.