International organisation for standardisation organisation internationale de normalisation
Yüklə 8,24 Mb.
|
5.8Non-deblocking loop filters5.8.1Adaptive loop filterA BoG was requested to discuss contributions in this area (H0409, H0094, H0420, H0483, H0372, H0277, H0418, H0571, H0413), coordinated by T. Yamakage. 5.8.1.1.1.1.1.1.1JCTVC-H0732 BoG report on Adaptive Loop Filtering [T. Yamakage] Break-out group meetings on Adaptive Loop Filtering were held at 14:30-18:00 on Feb. 5, 2012 and at 16:10-17:30 on Feb. 6, 2012. All BoG discussions were chaired by T. Yamakage. Approximately, 20 participants were present (including the chair). The BoG reviewed and summarized the non-CE proposals on Adaptive Loop Filtering (JCTVC-H0372, JCTVC-H0571, JCTVC-H0418, JCTVC-H0409, JCTVC-H0415, JCTVC-H0094, JCTVC-H0483, JCTVC-H0420). The BoG recommended adoption of JCTVC-H0409 and JCTVC-H0483. The BoG recommended further study of JCTVC-H0372, including subjective viewing in a CE or AHG. This was agreed. The BoG recommended further study of JCTVC-H0094 in a CE or AHG. This was agreed. The BoG found no evidence to revert the decision of Session A to adopt JCTVC-H0068. 5.8.1.1.1.1.1.1.2JCTVC-H0735 Analysis of differences of various ALF signalling schemes [T. Yamakage] This contribution discussed where to signal ALF information. It was noted that the current Wiener-based filter design technique for ALF would not be the only way to design the filter control parameters, and that enabling optimization over larger regions improves the opportunity for optimization while increasing latency. Control parameters can be placed in the APS, SH, or SD. When frame-level (or multi-frame) optimization is performed, it would be an undesirable burden to require the encoder to go back and CABAC-encode the control parameters within the slice data stream. For low-delay encoding, there are two possibly reasonable approaches:
The current WD has some syntax operation at a finer level than the LCU level. The value of this finer-than-LCU level of syntax effect is unknown. Class A gain for ALF is reported to be approximately 4.5% for the two 8 b sequences, and 5.7% for the two 10 b sequences. For RA HE, it is reportedly 5.6% for all 4 sequences coded with 8b encoding (not concentrated in any particular sequence). H0066 always has control at LCU level and also proposes signal processing simplifications. It was suggested that this may have some penalty, e.g. for small LCU operation. The H0274 "two-stage design" variant has control at either the LCU level or coarser (quadtree based) without signal processing changes to HM 5 (with a simplification to remove the RA mode). In principle, if the control is at the LCU level, it could be interleaved with the SD. With regard to removal of the BA or RA mode, another participant suggested a preference for removal of the BA mode rather than removal of the RA mode. Some other participants remarked that BA is better than RA mode. Decision: Adopt H0274 "two-stage design" with control at LCU level or coarser and simplification to remove the RA mode. Further study of additional simplification (e.g. per H0066) and BA vs. RA and placement of syntax is to be done in CE or AHG. 5.8.1.1.1.1.1.1.3JCTVC-H0094 Non-CE8: Modification of directional classification for block-based adaptation on ALF [M. Matsumura, S. Takamura, A. Shimizu, H. Jozawa (NTT)] The BoG (see H0732 BoG report) recommended further study of JCTVC-H0094 in a CE or AHG. This was encouraged. 5.8.1.1.1.1.1.1.4JCTVC-H0158 Non-CE8: Cross-check of modified directional classification for block-based adaptation on ALF (JCTVC-H0094) [T. Yoshino, K. Kawamura, S. Naito (KDDI)] 5.8.1.1.1.1.1.1.5JCTVC-H0136 Non-CE8: Report on visual artefacts by ALF virtual boundary processing and proposal on the filtering process modification [O. Nakagami, M. Ikeda, T. Suzuki (Sony)] This contribution reports that visual artefacts may occur that have a relation to ALF virtual boundary processing. The fixed position of the artefact pattern makes the artefact visible. However, the particular described problem is no longer a concern due to a change of the ALF filter shape. It seems useful to consider the potential for such an issue. It was commented that we had a discontinuity in the processing at a boundary, introduced by the ALF virtual boundary processing, and some concern was expressed that it might still cause some problem. It was noted that ALF is otherwise beneficial to the subjective quality of this sequence, so disabling it would not be desirable as a solution. 5.8.1.1.1.1.1.1.6JCTVC-H0163 Cross-check of JCTVC-H136 [A. Ichigaya, S. Iwamura (NHK)] [late]
The BoG (see H0732 BoG report) recommended further study of JCTVC-H0372, including subjective viewing in a CE or AHG. This was encouraged. 5.8.1.1.1.1.1.1.8JCTVC-H0661 Non-CE8: Crosscheck for Panasonic line memory reduction fix at independent slice and tile boundaries in JCTVC-H0372 [C.-Y. Chen, Y.-W. Huang (MediaTek)] [late]
This contribution seemed no longer important to consider as-is, due to the adopted change of filter shape. However, it was suggested to have some relevance with the modified filter shape. 5.8.1.1.1.1.1.1.10JCTVC-H0683 Non-CE8: Cross check of Modification of ALF BA classification [I. S. Chong, M. Karczewicz (Qualcomm)] [late]
It was asserted that, currently, loop filtering across slices (for all three filters) is controlled by a flag at the SPS level. See notes on H0391. 5.8.1.1.1.1.1.1.17JCTVC-H0409 Harmonization of number of ALF classes between RA and BA modes [E. Maani, A. Tabatabai (Sony)] 16 classes are proposed to be possible for the BA mode instead of the current 15 classes in WD5, in order to have the same number of classes between the RA and BA modes. This method merges the smooth areas into one class since they do not have substantial directional characteristics. The potential need for a more substantial cleanup than this (e.g. to use the same classes for both the RA and BA modes rather than just harmonizing the number of classes) was suggested and discussed. However, based on the information currently available, this seemed to be an improvement relative to our current design. Decision: Adopted as recommended by BoG (see H0732 BoG report). This was further discussed, and reconfirmed. Further study in a CE or AHG was encouraged to consider reducing the number of classes. 5.8.1.1.1.1.1.1.18JCTVC-H0674 Non-CE8: Cross-check of H0409 on Harmonization of number of ALF classes between RA and BA modes [P.Lai, F. Fernandes (Samsung)] [late 02-01] 5.8.1.1.1.1.1.1.19JCTVC-H0483 Harmonization of ALF Luma and Chroma Center Coefficient Prediction [J. Zhao, A. Segall (Sharp)] This document reports an inconsistency in the design of the ALF coefficient coding for luma and chroma. Specifically, the prediction method employed for luma coefficients is the negative of the prediction method used for chroma coefficients. This document proposed a modification that was asserted to better harmonize the luma and chroma center coefficient prediction methods. Decision: Adopted as recommended by BoG (see H0732 BoG report). 5.8.1.1.1.1.1.1.20JCTVC-H0571 ALF with multi shapes versus ALF with single shape [Faouzi Kossentini, Hsan Guermazi (eBrisk Video Inc.), Wang Lai, F. Fernandes (Samsung)] Yüklə 8,24 Mb. Dostları ilə paylaş:
|