Organisation internationale de normalisation

4.3Discussion and Conclusions

O. Au was requested to coordinate collection of cross-verification summary information for JCTVC-D167 and JCTVC-D295, as checked in JCTVC-D098, JCTVC-D169, and JCTVC-D247. The information collected is recorded above.

There was some discussion of creating a unification proposal relating to JCTVC-D167, JCTVC-D295 and JCTVC-D098. The following contribution was then submitted.

4.3.1.1.1.1.1.1.1JCTVC-D448 CE1 Subtest1: A joint proposal of candidate-based decoder-side motion vector derivation [Y.-J. Chiu, L.Xu, W. Zhang, H. Jiang (Intel), M. Yang, S. Lin, H. Yu (Huawei)] (late registration Tuesday 25th after start of meeting, uploaded Tuesday 25th, sixth day of meeting)

This contribution described a joint recommendation of the techniques of JCTVC-D167 and JCTVC-D295: a candidate-based Decoder-side Motion Vector Derivation (DMVD) technique to improve the coding efficiency relative to TMuC 0.9 reference software. The recommendation of the Candidate-based-DMVD (C-DMVD) had reportedly demonstrated an overall 2.0% BD BR improvement for the RA HE configuration under the common test conditions with 23% increase in encoding time and 5% increase in decoding time. A description about the normative syntax change about C-DMVD in the form of HEVC specification was included. Mitsubishi cross-checked the result and the estimate of the increase in encoding time was around 20% and the increase of decoding time was around 9%.

The range of per-sequence BD BR improvement was reported as 0.9-3.4%.

This proposal is only for the RA case.

It was remarked that the described scheme creates a dependency between the the parsing process and the reconstructed sample values of the previously-decoded reference pictures, and removing this dependency might reportedly degrade the performance by about 0.7%.

This is somewhat different than having a dependency on the motion vector values in the previously-decoded reference pictures. It is also different from having a dependency on the reconstructed sample values of the current picture (which could perhaps provide another 0.5% gain ‒ e.g., see JCTVC-D295).

It was noted that the proposal seems to increase memory bandwidth for the motion compensation process by up to 9x relative to ordinary bi-prediction. Decoder bandwidth is a major concern in the design.

Due to concern about the effect of this proposal on the decoder memory bandwidth, it was concluded that further study was needed. Such further study was encouraged, to try to find an alternative approach with a considerable reduction in memory bandwidth.

It was suggested that limiting the size of the fetch area might be beneficial.

4.3.1.1.1.1.1.1.2JCTVC-D450 Cross-check report on harmonized DMVD technique (JCTVC-D448) [Shun-ichi Sekiguchi, Kazuo Sugimoto(Mitsubishi)] (late registration Tuesday 25th after start of meeting, uploaded Thursday 27th, near the end of the meeting)

Cross-check of JCTVC-D448 (without reported difficulty).