Organisation internationale de normalisation

12.3Discussion and Conclusions

12.3.1First part of CE9: Study of the configuration of the Skip, Direct and Merge modes

In summary, it was reported that PU merge seems generally better than CU merge.

Disabling direct mode seems to provide improvements, as this has some redundancy with PU merge.

Combination 3.1.e and 3.1.t seem to have the best overall compression efficiency – with approximately 1% and 1.2% improvement overall for these two. Both of these were indicated, by runtime measures, to actually reduce complexity relative to the anchors.

It was remarked that PU merge may be difficult to use, especially in hardware, due to a difficulty in parallelization for that feature. Another participant commented that this is also true for other schemes as well, such as direct mode (at least for spatial direct mode).

A participant that had contributed JCTVC-D233 ("3.1.u" above, which was not actually part of the original CE plan) remarked that this contribution is closely related and takes into account some of the conclusions of this work and provides further gain and complexity reduction. That contribution reportedly also removed the redundancy in splitting of CUs and PUs.

A participant remarked that JCTVC-D370 should perhaps be considered in relation to this, as it contains an improvement of LCEC entropy coding. Another participant, somewhat similarly, remarked that it may be best to pay the most attention to the results from the CABAC entropy coder, under the assumption that the other entropy coder would eventually be improved to more closely match its performance. A participant indicated that the basic relative results would be unlikely to change if only the CABAC cases were considered.

A participant remarked that a parsing problem exists in the use of temporal MV prediction in the current HM design and that there is at least one contribution JCTVC-D197 proposing to fix that. A participant remarked that all of the best performing proposals have temporal MV prediction, and that a similar fix could probably be applied to all or most of them without substantial impact on their relative performance to each other.

At this point, the JCTVC-D314 "3.1.e" and JCTVC-D411 "3.1.t" combinations appeared to be the best among what was planned to be tested in the CE. Those and JCTVC-D233 were then further reviewed (possibly related other contributions) before making a decision.

In regard to JCTVC-D233: Several aspects of the document relate to JCTVC-D314 "3.1.e". Section 3 of the document is a new proposal to remove redundancy between CU and PU splitting. Two variants were proposed.

The first variant, without asymmetric motion partitioning (AMP), reportedly showed, relative to JCTVC-D314 "3.1.e", about 0.1% reduction of coding efficiency with a substantial reduction of encoder runtime (roughly 25%).

A participant remarked that it may be possible to achieve the same degree of encoder simplification without modification of the syntax, semantics, or decoding process. The participant also remarked that a different syntax modification may provide a simpler way to accomplish a similar result. It was suggested that experiment results for these alternative approaches could become available during the meeting.

There was a suggestion to decompose this into two elements, initially focusing on the relationship with "3.1.e". The group then discussed these elements as follows;

• 
  Removal of NxN split of CU into 4 PUs for inter cases except for the smallest CU size. It was noted that JCTVC-D314, JCTVC-D087 and JCTVC-D356 all advocated this same change.

  Decision: It was agreed that this should be done.

  
  
• 
  Replacing 2NxN and Nx2N PU splits with new "partial merge" modes, which reduces the signaling information for this selection. We revisited this aspect later during the meeting.
  

It was remarked that removing the NxN PU split for the smallest CU size also may be good to do. Further study on that topic was suggested (e.g., in a CE).

Further discussion of "3.1.t" versus "3.1.e":

The difference between "3.1.e" and "3.1.t" is that in "3.1.t" a skip mode is enabled with derivation of the MV from MVC, whereas in "3.1.e", there is no skip mode but there is a merge with no residual (which the encoder currently selects only if the residual coding results in no residual).

A participant remarked that "3.1.t" should be used for reasons of draft stability, and another participant thought that "3.1.t" may be better for encoder parallelism.

The second variant, with the asymmetric motion partitioning (AMP) feature added, reportedly showed, relative to JCTVC-D314 "3.1.e", about 0.7% improvement of coding efficiency with some increase of encoder runtime (roughly 9%).

It was remarked that the prior asymmetric motion partitioning (AMP) feature was not included in the WD, and this variant adds it back in some way – and it may be desirable to compare this proposed method for that feature with the prior proposed method for that feature.

BoG activity for further discussion was requested, coordinated by T.K. Tan & W.-J. Han. This resulted in the submission of JCTVC-D441. Please refer to the notes on that report for additional conclusions.