Organisation internationale de normalisation
Subset 2: Wiener-based in-loop filters
Yüklə 7,54 Mb.
|
11.2.2Subset 2: Wiener-based in-loop filters11.2.2.1.1.1.1.1.1JCTVC-D119 CE8 Subset2: A Joint Proposal on Improving the Adaptive Loop Filter in TMuC0.9 by MediaTek, Qualcomm, and Toshiba [C.-Y. Chen, C.-M. Fu, C.-Y. Tsai, Y.-W. Huang, S. Lei, M. Karczewicz, I. S. Chong, T. Yamakage, T. Chujoh, T. Watanabe] New joint proposal by three of the (five) companies that participated in the CE (MediaTek, Qualcomm, and Toshiba). For Subset 2, JCTVC-D119 was a new proposal from MediaTek, Qualcomm, and Toshiba. It differs from what was planned to be tested in this CE. A participant remarked that this contribution includes a combination of different techniques that would be desirable to assess individually. It was remarked that the chroma filtering should be investigated. It was noted that although complexity reduction was a goal of the work, the worst case complexity is not reduced by the proposals. It was noted that JCTVC-D119 proposes an ability for the encoder to add an offset as well as to apply filtering. The adaptive offset method of JCTVC-D122 was integrated. However it was not operated in series with ALF, but rather was made switchable. An additional hierarchy "filter unit" (FU) is introduced (always rectangular) – this could be in conflict with the case of multiple (non-rectangular) slices. Elements of the proposed design were as follows:
NOTE: Statistics are collected during the DF pass, then decisions are made, and one pass is performed for application of the ALF.
In the current design, there is no FU level – the filter is controlled at the CU level. It was remarked that there is an issue in regard to applying the FU concept to slice-structured coding that would need to be resolved. In JCTVC-D122, the "adaptive offset" technique could be used in addition to another filter instead of as an alternative as proposed in this contribution. This provides more reported gain, but adds extra processing. Overall results were reported as follows: HE AI 0.1% worse, HE RA 0.4% better, HD LD 1.9% better in BD rate average. Decoding time reduced by 15%, encoding time approximately unchanged. It was remarked that the number of encoding passes (and ALF overall) does not affect our total software simulation runtimes by much (although it would be a serious problem for implementations). The presenter indicated that the one-pass encoding would cause a loss of fidelity by about 1%, if the other aspects were not included. It was noted that non-normative configuration changes will be made to the test conditions, which could affect signal propagation fidelity in the LD predictive case. The presenter asserted that the changes proposed would not conflict with those of the other proposals in the CE. It was remarked that there is a relationship between the deblocking filter behavior and ALF behavior. It was remarked that the subjective gain from ALF is much more substantial than the objective gain. It was remarked that this proposes an increase in the complexity of the ALF design. A participant asked what the impact would be of just using LCU signaling instead of introducing an additional FU layer. A proponent indicated that this would probably increase the amount of overhead to an unacceptable degree. Since there are several distinct modifications proposed here, some of which seem to substantially increase ALF complexity, and since there are other proposals to consider that are closely related, further study would be needed to determine the appropriate action in regard to this proposal. It was suggested (despite the reported 1% loss) that it would be nice to enable a way to configure the software to not use the current highly-multipass (16 pass) encoder behavior (while retaining the current behavior when configured for multi-pass). It may also be nice to be able to configure a multi-pass usage between the extremes of single-pass and 16 pass. Decision: This was agreed. The software coordinator is generally delegated with the ability to decide when additional similar sorts of (non-normative) configurability are desirable. 11.2.2.1.1.1.1.1.2JCTVC-D216 CE8: Cross-check results of the adaptive loop filter of MediaTek, Qualcomm and Toshiba [Semih Esenlik, Matthias Narroschke] Cross-check of JCTVC-D119. 11.2.2.1.1.1.1.1.3JCTVC-D114 CE8: Cross-check result of MQT adaptive loop filter [Tomohiro Ikai] Cross-check of JCTVC-D119. 11.2.2.1.1.1.1.1.4JCTVC-D170 CE8: Cross-check of MQT_ALF results from MediaTek, Qualcomm and Toshiba (JCTVC-D119) [Y.-J. Chiu, L. Xu, W. Zhang, H. Jiang (Intel)] Cross-check of JCTVC-D119. 11.2.2.1.1.1.1.1.5JCTVC-D115 CE8: DF-combined adaptive loop filter [Tomohiro Ikai, T. Yamazaki (Sharp)] This contribution was a CE8 subtest 2 (Wiener-based in-loop filters) proposal. This proposes a two-input filter, where one input is the output of the DF and the other is the pre-DF reconstruction. The proposed method does not use sample-by-sample adaptation of the filter. It was remarked that such adaptation could be used and would provide some gain if used – perhaps somewhat more than 1%. (And a similar amount of gain would be lost if the adaptation were removed in the current design.) The spatial extent of the proposed filtering supports diamond kernels of shape 7x7, 9x9, and 11x11. It was remarked that adding 11x11 to the current design would provide an improvement, but that this had not been done for reasons of complexity. It was remarked that this proposal tries to align luma and chroma processing elements. In our current reference configuration, chroma uses square filters while luma uses diamond ones. The proposed technique proposes the ability to process the two inputs in parallel. The experimental results for HE reportedly show 0.2%, 0.2%, and 0.7% BR reduction IO, RA, and LD. The decoding time ratio is 85% and the encoding runtime ratio is somewhat reduced. Additional experimental results show the DF-combined improvement, which is the bdrate difference between the proposal's two input result and a similar design (not the current ALF) using one input is 0.9%, 0.7%, 0.7% in IO, RA, and LD respectively. Basically three aspects:
A participant remarked that subjective viewing results from viewing conducted for TE10 at the Guangzhou meeting demonstrated significant visual subjective improvement from the sample-by-sample adaptivity aspect of the current ALF (based on comparison between a Toshiba proposal without that adaptivity vs. QCALF and a Panasonic proposal). In a sense, this approach can be roughly interpreted as a special case of the JCTVC-D217 three-input approach. Further study was encouraged. 11.2.2.1.1.1.1.1.6JCTVC-D215 CE8: Cross-check results of the DF-combined adaptive loop filter of Sharp [Matthias Narroschke, Semih Esenlik] Cross-check of JCTVC-D115. 11.2.2.1.1.1.1.1.7JCTVC-D211 CE8 Subset2: Verification results of Sharp's Proposal JCTVC-D115 [T. Yamakage (Toshiba)] Cross-check of JCTVC-D115. 11.2.2.1.1.1.1.1.8JCTVC-D217 CE 8: Results for adaptive loop filter using prediction and residual (3-Input-ALF) [S. Esenlik, M. Narroschke, T. Wedi (Panasonic)] This contribution presented results of the Adaptive loop filter using both the prediction and residual (3-Input-ALF). An average Y-BD BR gain of 0.9% for AI HE, 0.8% for RA HE, and 0.4% for LD HE, compared to the official anchors can reportedly be achieved. The proposal produces an output that is a weighted combination of:
The decoder (if it wants to use this feature) would need additional line buffers to store the extra signals. The encoder may need multipass operation (for reference frames) with full frame stores for these extra signals. At the CU level, signaling specifies whether to use the weighted combination or just the output of the DF (as in the current HM design). The same weighting values are applied for all regions of the picture for which the weighting is enabled. It was noted that one example set of values for this weighted combination would entirely disable the DF. In (informal) testing done at the last meeting, it was reported that there was no (significant) visual degradation observed in any cases, and there seemed to be some improvement for one sequence. It was remarked that other DF and ALF improvements might obviate the need for the three inputs. Further study was encouraged. 11.2.2.1.1.1.1.1.9JCTVC-D237 CE8 Subset2: Verification results of Panasonic's Proposal JCTVC-D217 [T. Yamakage (Toshiba), I.S. Chong (Qualcomm), Y.W. Huang (MediaTek)] Cross-check of JCTVC-D217. The software was examined and determined to be consistent with the proposed technique and to be a straightforward implementation of the described scheme. Yüklə 7,54 Mb. Dostları ilə paylaş:
|