Joint Collaborative Team on Video Coding (jct-vc) of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

5.14Entropy coding

JCTVC-I0065 Modification on CABAC initialization [K. Sugimoto, A. Minezawa, S. Sekiguchi, T. Murakami (Mitsubishi)]

This contribution proposes a modification on initialization table selection for CABAC context. The proposed method uses two initialization tables instead of three tables as in HM6. In P-slices and B-slices, initializationType = 0 and 1 can be selected instead of 1 and 2 in HM6, while initializationType=0 is used for I-slices same as in HM6. To verify the performance, sequences with frequent scene changes are used with applying initializationType = 0 in slices just after scene change. Simulation results reportedly show that performance gain is 0.1%, 0.1%, 0.1%, 0.1%, 0.1%, and 0.1% for RA-main, LB-main, LP-main, RA-HE10, LB-HE10 and LP-HE10 settings respectively using HM-6.0 in multiple 1500 byte slice mode with QP = 12, 17, 22 and 27. It is reported that the BD BR gain with QP = 32, 37, 42 and 47 is 0.1%, 0.2%, 0.1%, 0.1%, 0.1%, and 0.2% respectively.

One motivation is to allow a P or B slice at a scene change to use the intra-oriented table. The gain would bigger in that case.

Sequences outside the test set seemed potentially valuable to test.

The conditions under which this was tested seemed a bit irregular.

It was also suggested to check whether alternative values for the P and B slice init tables would provide some benefit.

Testing on small slices, perhaps with the encoding rule to carry over the selection from the best fit for the preceding slice, might be beneficial. Testing like this had been done historically for AVC.

The intra tables don't include initialization values for some syntax elements. In such cases, the contributor just copied the initialization values from another table that had them.

At the moment we do not seem confident of what we currently have for the initialization, and believe that further study (AHG or CE) is needed.

JCTVC-I0204 Reduction in context coded bins for ref_idx and cu_qp_delta [V. Sze, M. Budagavi (TI)]

To improve the throughput of the CABAC, it was asserted that the number of context coded bins should be reduced. Hardware is designed for the worst case, thus it is also important to reduce the worst case number of context coded bins. In HM-6.0, ref_idx and cu_qp_delta reportedly account for 28% of the context coded bins in the worst case. This is due to the fact that they are binarized with truncated unary and all the bins are context coded. This contribution proposes using a combination of truncated unary and Exp-Golomb coding (TUEGk), similar to what is used for motion vector difference. Only the truncated unary portion is context coded, and the Exp-Golomb portion is bypass coded. This was implemented in HM-6.0 for various combinations of cMax and EGk. With TU cMax=4 and EG0, no coding loss was observed under common conditions and the worst case number of context coded bins is reduced by over 20%. Since cu_qp_delta is not signalled under common conditions, it was testedin HM-6.1 under the conditions specified by CE4 on quantization. There is reportedly negligible coding efficiency impact for cMax=4 and EG0 (between -0.02 to 0.03).

Although the binarization was changed, the initialization was not.

It was noted that the severity of the impact of the worst case could be limited by profile constraints on the spatial granularity of QP changes.

It was remarked that the threshold of where to stop using context coding in the binarization might depend on the way the scheme was tested.

Curiosity was also expressed regarding the worst case for syntax elements of SAO and ALF.

Uses bypass coding after the 4th bin.

The reference index part is closely related to I0352. See notes for I0594.

JCTVC-I0544 Cross-check of Reduction in context coded bins for ref_idx and cu_qp_delta (JCTVC-I0204) [J. Wang, X. Yu, D. He (RIM)] [late]
JCTVC-I0395 Cross check of JCTVC-I0204 on reduction in context coded bins for ref_idx and cu_qp_delta [J. Xu, C. Auyeung (Sony)] [late]
JCTVC-I0352 Bypass bins for reference index coding [V. Seregin, X. Wang, J. Chen, M. Karczewicz (Qualcomm)]
Addressing essentially the same problem as the reference index part of I0204, but in a somewhat different way. Does not change binarization, but switches to bypass coding after the 2nd bin. 8% of the reference index bins in the common conditions then used bypass coding. No significant coding efficiency impact was reported. See notes for I0594.

JCTVC-I0594 Bypass bins for reference index and delta QP coding [V. Seregin, X. Wang, W.-J. Chien, J. Chen, M. Karczewicz (Qualcomm), V. Sze, M. Budagavi (TI)] [late]

This contribution asserts that the worst case number of context coded bins for ref_idx and cu_qp_delta should be reduced. Modification of reference index coding is proposed according to which all bins after the second bin are coded with CABAC bypass mode as proposed in JCTVC-I0352. Experimental results reportedly show no performance change and that 8% of the affected bins for these syntax elements are coded using bypass bins on average under four common test configurations. The proposal also results in one context removal. For delta QP coding, it proposes using a combination of truncated unary and Exp-Golomb binarization (TU4EG0) as proposed in JCTVC-I0204, similar to what is used for motion vector differences . Since delta QP is not signalled under common conditions, it was tested in HM-6.1 under the conditions specified by CE4 on Quantization. There was reportedly a negligible coding efficiency impact (between −0.02 to 0.03) from this aspect.

This contribution was submitted as a proposal to resolve the differences between I0204 and I0352.

The rationale for choosing the I0352 approach for reference index coding was to choose the one with the better worst-case complexity behaviour. (This could have a hypothetical R-D penalty in some non-CTC usage.)

For further study (AHG with mandate to study reduction of context-coded bins in general), including a similar issue for SAO (and perhaps for transform coefficient coding).

JCTVC-I0446 Crosscheck of bypass bins for reference index coding in JCTVC-I0352 [Y.-W. Chen, Y.-W. Huang (MediaTek)] [late]
JCTVC-I0276 Updated design of the probability estimator and the coding engine of the M Coder [J. Stegemann, H. Kirchhoffer, D. Marpe, T. Wiegand (Fraunhofer HHI)]

This contribution proposes a set of modifications to the current design of the probability estimator and the entropy coding engine of HEVC. The state-based probability estimator is extended from 63 states to 112 states and the table for the interval subdivision of the arithmetic coding engine is adjusted to conform with the internal state of the probability models. While each probability model requires 1 additional bit of memory the average gains in coding efficiency reportedly range from 0.3 to 0.5%.

Currently we have the same probability estimator for HEVC as for AVC.

This seems to be primarily a coding efficiency proposal.

Some tables are made larger, and some smaller (range table).

There were some significant differences between the first version of the proposal and the second version uploaded 10 days later.

It was remarked that consistency with AVC may have value, and significant effort has been put into CABAC over the past decade.

The stability of the design is also important.

However, it may be valuable to encourage study of the implications of the proposed modification.

For further study.

JCTVC-I0509 Cross-check of updated design of the probability estimator and the coding engine of the M Coder in JCTVC-I0276 [?? (??)] [late]
JCTVC-I0299 Experimental results for Multi-Parameter Probability Update in HM6.0 [A. Alshin, E. Alshina, J.-H. Park (Samsung)] [late]

Multi-parameter probability update method is an alternative probability estimation method based on combination of multiple estimations by different models. Experimental results for HM6.0 are presented in this proposal. The proposed technique reportedly shows 0.6% Luma BD-rateBD-BR gain (and 0.2% Chroma BD-rateBD-BR gain). It was reported that the proposed technique can be effectively realized in integer arithmetic.

No text or software was provided, and no cross-check was available. It was remarked that it has a substantial memory capacity requirement.

This did not appear adequately mature for action, aside from concerns regarding stability.

No action was taken on this.