Organisation internationale de normalisation

Test 6 – Combination of test 1 and the current IBC design (JCTVC-S0081)

In this test, a CU-level flag is sent to indicate IBC flag. Two sub-tests were conducted. In test 6a, both IBC merge mode and IBC AMVP mode are unified with inter mode. In test 6b, only IBC merge mode is unified with inter mode, i.e., BVD and BVP coding are the same as the current IBC design.

From discussion: Basically, test 6a/b are equivalent with 5a/b except that the IBC flag is signalled at the CU level, whereas test 5x uses PU level signalling. Further, 6a used the BV difference coding from SCM, whereas 5a uses the conventional MV difference coding.

The summarized results of test 6a are shown below –

The summarized results of test 6b are shown below –

From the discussion:

Generally, a significant compression benefit is observed by adding skip/merge to IBC. The method from test 3 does this in a IBC specific way (to some extend modifying the current merge list construction of MV coding), whereas the methods of test 1, 5, and 6 are using the same skip/merge process as used in MV coding.

1. 
   Some of the methods (test1, test 5a) would have the same worst-case-context-coded-bin problem as discussed in context of CE1, since they are using the MVD coding for IBC vectors.
   
2. 
   It is also questioned whether invoking merge at 4x4 blocks could cause some complexity problems.
   

1. and 2. should be in detail investigated in the BoG on complexity assessment of IBC vector coding (C. Pang, S. Liu)

See further notes under S0294 – to be further investigated in ongoing CE.

The PU-level switching gives some gain (both from test 2 and comparing test 5b against 6b, which should otherwise be identical), in the range of 1.5% for TGwM class, around 0.3% for some other classes. It is however reported in the cross-check that part of this gain may be due to the fact that the RD check for 2NxN and Nx2N partitions is disabled in the default SCM settings at 16x16 CU, whereas it is enabled for the PU level method (note: those partitions are already supported by current IBC).

One expert reports that according to his experience (in the context of test 3) the gain by allowing 8x16 and 16x8 RD check could give around 0.8%, which is doubted by yet another expert.

Additional results were asked to be provided reporting the gain of PU level switching against a modified SCM reference that enables 16x8 and 8x16 RD checks.

It was later reported that the gain drops by rougly 0.5% - to be further investigated in ongoing CE, but then with additional enabling the RD checks. (does not have highest, but could be interesting to give more insight when comparing CU-level IBC signalling versus the refidx based approach)

See notes on BoG S0294.

13.0.0.1.1.1.1.1.78JCTVC-S0022 CE2: Summary report for Core Experiment 2 on intra block copy signalling and partitioning [J. Xu, S. Liu, K. Rapaka, X. Xiu (CE coordinators)]

13.0.0.1.1.1.1.1.79JCTVC-S0294 BoG report on complexity assessment of IBC block vector coding [C. Pang, S. Liu]

(Chaired by JRO & GJS on Thursday 10-23 a.m.)

A BoG was held on Saturday October 18, 2014, 9:00 am to 11:15 am to provide more detailed complexity assessment of IBC vector coding. The agreed mandates of the BoG were:

1. 
   Analyze the worst case number of context coded bins in IBC for the current SCM and the methods CE1 Test 1.1 (JCTVC-S0142) and Test 1.4 (JCTVC-S0162). Compare against current 4x8 MV coding, including considerations about reference index and mode coding.
   
2. 
   Discuss whether invoking merge at 4x4 blocks could cause some complexity problems.
   

Item 1, intra case: