International organisation for standardisation
Parallel ME (merge/skip) level 2N x 2N
Yüklə 9,08 Mb.
|
- Bu səhifədəki naviqasiya:
- Parallel ME level 2N x 2N (LCU = 64x64) RA-HE (%) RA-LC (%)
This is better than the loss for the same encoder structure without the proposal, reported as:
It was remarked that the merge/skip design is not really finalized yet, so that even if we might like this, it may be an over-optimization to a particular instantiation of a moving target at this stage. This (or something like it) may be desirable to keep in mind for consideration in the future work. It was suggested to include this in a CE. 6.9.1.1.1.1.1.1.2JCTVC-F208 Cross-check of parallelized merge/skip mode (JCTVC-F069) [B. Li (USTC), J. Xu (Microsoft)] 6.9.1.1.1.1.1.1.3JCTVC-F465 Item 5 (Partial merge) of Experiments on tools in Working Daft (WD) and HEVC Test Mode (HM-3.0) [I.-K. Kim, E. Alshina, J. Chen, T. Lee, W.-J. Han, J. H. Park (Samsung), V. Sze, M. Budagavi, M. Zhou (TI)] Item 5: Partial merge. Has been taken care of (JCTVC-F082). 6.9.1.1.1.1.1.1.4JCTVC-F325 Modified temporal MV derivation process for merge/skip mode [T. Shiodera, A. Tanizawa, T. Chujoh, T. Yamakage (Toshiba)] This contribution proposed a modified derivation process of temporal merge/skip mode in B-slice and presents associated experimental results. The temporal merge/skip mode was introduced in latest HM test model. This can reportedly reduce an overhead of motion information by reusing temporal motion information in the current prediction unit. However, when two kinds of motion information used for bi-prediction refer same reference block, it is noted that that the prediction value for bi-prediction is identical to one for uni-prediction. In this contribution, two modifications for derivation process of temporal merge/skip mode in B-slice are proposed to be introduced in order to improve both complexity and coding efficiency. The first modification is to change inter prediction processing from bi-prediction to uni-prediction in order to reduce the computational complexity of the redundant motion compensation process. The other one is to change the derivation process of list 1 motion information for temporal merge/skip mode in B-slice, in order to avoid deriving same motion information in the current PU. This modification can reportedly improve coding efficiency. Experimental results reportedly show that the proposed methods can achieve an average luma BD-rate gain of 0.2% and 0.3% for the LB-HE and LB-LC cases, respectively. Average decoding times are reportedly 98% and 96% for LB-HE and LB-LC cases, compared with the common condition anchor, respectively. The first aspect is something the decoder might be able to do on its own, by simply recognizing that bi-prediction will obtain the same results as uni-prediction and is therefore unnecessary to be applied. This could perhaps be a desirable thing to include in our reference software if suitable source code is provided that does not harm the readability of the software – the decision on that is delegated to the software coordinator and software development AHG. The second aspect provides a small amount of coding gain (0.2-0.3%) in the LB case. It was remarked that there were some bugs in the relevant part of the version of the software that was used, so we may not be able to rely on these results. Further study was encouraged to determine how this may work in the HM4 context. 6.9.1.1.1.1.1.1.5JCTVC-F356 Motion compensation complexity reduction for bi-prediction [H. Y. Kim (ETRI), K. Y. Kim, G. H. Park (KHU), S.-C. Lim, J. Lee, J. S. Choi (ETRI)] This contribution was closely related to JCTVC-F325. This contribution reports that roughly 30% and 5% of the areas of forward B-Slices are observed to have identical motion information within each PU, under HM3 LD and RA configuration, respectively. It was indicated that when the L0 and L1 motion information is the same, the L1 interpolation process and the weighted averaging process in HM3 could be bypassed for complexity reduction. The idea of integrating the decoder-only optimization scheme into the reference software, as described in the section discussing JCTVC-F325, was also relevant to this contribution. Results of this technique were provided in this contribution as "method A" (with a reported decoding time savings around 4% in the LB case). Several slightly different methods were presented for handling the same-motion cases. A significant percentage of predictions in the LB case are bipredictions that can degenerate to unipredictions if recognized. As a decoder-only non-normative trick, the decoder can take advantage of this. We should try to ensure in the future that the prediction generated by uniprediction will be the same as the prediction generated by biprediction with identical parameters. 6.9.1.1.1.1.1.1.6JCTVC-F712 Additional results on JCTVC-F356 (MC complexity reduction) [H. Y. Kim (ETRI), K. Y. Kim, G. H. Park (KHU), S.-C. Lim, J. Lee, J. S. Choi (ETRI)] [late reg. 07-11, upload 07-13] This contribution provides two additional results related to JCTVC-F356 (Motion compensation complexity reduction for bi-prediction). The goal of this is to improve coding efficiency when the collocated motion information of a bi-predicted Merge/SKIP PU are identical. In Method-A, in this case the neighboring PUs are searched for a non-zero motion vector. If such an MV is found, it is used as mvL1Col. In Method-B, Method-A and Method-2 of JCTVC-F356 are combined for motion compensation complexity reduction. The combined method reportedly achieves average coding gain of 0.3% and 0.5% for LD-HE and LD-LC configurations, respectively. A participant suggested to, instead of looking for an MV that is zero, look for one that is different from what would otherwise be inferred. This could potentially be considered as part of a continuation of CE9. 6.9.1.1.1.1.1.1.7JCTVC-F769 Cross Check of JCTVC-F712 [D. Flynn (BBC)] [late reg. 07-20, upload 07-20]
Yüklə 9,08 Mb. Dostları ilə paylaş:
|