International organisation for standardisation organisation internationale de normalisation

5.16.6CBF

No contributions noted.

5.16.7General/other transform coefficient coding contributions

5.16.7.1.1.1.1.1.1JCTVC-H0228 Non-CE11: Proposed Cleanup for Transform Coefficient Coding [T. Nguyen, D. Marpe, T. Wiegand (Fraunhofer HHI)]

This contributiono described a scheme for parameter selection related to the coding of transform coefficient levels with CABAC. The change relies on a subset of rules for context model selection and Rice parameter derivation. All transform block sizes and shapes would apply the proposed context model selection, Rice parameter derivation method, and use the same set of context models for the respective bin index. Significance coding and level coding were proposed to be substantially changed.

The last coefficient coding was unchanged. Characteristics of the proposal were summarized as follows:

• 
  Used the same subset of rules for context model selection and Rice parameter derivation
  
• 
  Used the same subset of rules for all transform block sizes and shapes
  
• 
  Used the same subset of context models for all transform block sizes and shapes
  
• 
  Enabled the pre-calculation of context models selected for next scan positions with respective Rice parameter
  

BR reduction (HE) was reportd as 0.7% AI, 0.5% RA, 0.2% LD.

It was asked whether anything changed in the scan, and this was indicated not to be the case – the scan would be the same as in the current design (no cleanup).

It was mentioned that various things are changed in the software which are not documented in the contribution. As one example, the init value for ALF was changed. Some concerns were expressed about the possibility of reduced throughput and parallelism due to larger amount of dependencies. For example, it would be necessary to completely decode certain coefficients before the next ones could be decoded.

The memory requirement seemed likely to be higher.

Further study (in a CE) was encouraged. It was suggested to investigate the benefit of the different elements and the impact on throughput.

5.16.7.1.1.1.1.1.2JCTVC-H0699 Non-CE11: Cross-check of proposal JCTVC-H0228 on transform coefficient coding [V. Sze (TI)] [late 02-03]

The cross checker indicated that this proposal introduces some additional dependencies and may provide lower throughput.

5.16.7.1.1.1.1.1.3JCTVC-H0399 Harmonization of coefficient coding [J. Xu, A. Tabatabai (Sony)]

This proposal would change coefficient coding for TUs with diagonal scan. The proposed solution reportedly had coding efficiency improvement compared to HM5.0. Experimental results reportedly showed that coding efficiency gain is 0.00% for AI HE, -0.13% for AI LC, -0.12% for RA HE, -0.24% for RA LC, -0.13% for LD HE, -0.30% for LD LC and 0.08% for RA HE10.

The proposal used a 4x4 sub-scan for 8x8 blocks as in H0284

For diagonal scan: a multi-level significance scan and context models from 16x16 were also applied in 4x4 and 8x8 cases.

It was asked what is the main reason for the reported gain, and this was reportedly most likely to be due to the neighbour-based context model.

This approach was supported by several experts as a desirable harmonization, and suggested that the BoG on significance map coding should review other proposals under this aspect.

Issues raised in BoG discussions included the following: the unification across diagonal scan implies that there is no unification on the 4x4 and 8x8 cases; as the vertical and horizontal scans remain position based. One expert stated a desire to to keep the position based model for 4x4 and 8x8: that there was no need for parallelization for a better throughput. One expert stated that a position based scheme is not simpler than a template scheme. One expert requested harmonization with MDCS and noted that it is important to clean-up that part of the software. (It was mentioned that JCTVC-H0379 could be related to this harmonization.) Concerns were expressed on the dependencies created in 8x8 when applying the 16x16 pattern. Some experts want to use only smaller TUs for the worst case scenario.

Further discussion in plenary was later conducted:

Decision: (following BoG recommendation) Adopt multi-level significance map coding and contexts reusing 16x16 and 32x32 cases for coefficient group flag for the 4x4 (sub)block scans in all block sizes.

Regarding the third item: using the same significant map coding (template mask based approach) for all diagonal scans, that is for 8x8 and 4x4 diagonal scans. Some gains were reported. It was recommended that this should be further studied in a CE.

It was remarked that in general it would be more desirable to have a full harmonization including the horizontal and vertical scans which are specific now for 4x4 and 8x8 cases and that this should become part of the CE (H0379).

5.16.7.1.1.1.1.1.4JCTVC-H0684 Cross-verification of JCTVC-H0399 on harmonization of coefficient coding [V. Seregin, J. Sole (Qualcomm)] [late 02-01]

The cross-verifier supported this as a desirable harmonization, and commented that it would be even better if the coding in horizontal/vertical scans would be further unified with the diagonal case.

5.16.7.1.1.1.1.1.5JCTVC-H0249 Non-CE11: Coefficient sub-sampling for large transform [Y. Chiu, W. Zhang, L. Xu, Y. Han (Intel)] [late]

This contribution presented a sub-sampling process to the transform coefficients with the aim to improve the hardware design bottleneck of the worst case of the maximum number of transform coefficients to be decoded. The sub-sampling operation is applied on the transform coefficients of the high frequency band, and it results in a reduction of the number of transform coefficients. Compared to the test anchor of HM5.0, one configuration of the proposed sub-sampling technique reportedly enables the worst case of the maximum number of transform coefficient to be reduced from 1024 to 768 for the 32x32 transform, and the worst case of the maximum number of transform coefficient to be reduced from 256 to 224 for 16x16 transform with an overall 0.1% BD bit rate change along with slight reduction in both encoding and decoding CPU run time. Subjective tests reportedly showed that no obvious visual quality degradation was introduced by the proposed coefficient sub-sampling.

It was commented that this can be done non-normatively by any encoder. There was no support by other experts, and no action taken.

5.16.7.1.1.1.1.1.6JCTVC-H0677 Non-CE11: Crosscheck for Intel Coefficient Sub-sampling for Large Transform in JCTVC-H0249 [Q. Huang, X. Guo (MediaTek)] [late]