Joint Collaborative Team on Video Coding (jct-vc) of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11
Plenary Discussions and BoG Reports
Yüklə 1,12 Mb.
|
6Plenary Discussions and BoG Reports6.1Project development6.2BoGsJCTVC-I0576 BoG report on SAO [Y.-W. Huang] Recommendations of BoG:
Decision: Adopt as stated in 4 bullets above – proponents are asked to provide a single text with integrated adopted SAO changes by Friday (Breakout: YW Huang) – this shall be provided with a detailed documentation which change came from which adoption, and if it is detected that further cleanup/correction is needed, it shall be reported. (In regard to I0066, it was later remarked offline that I0066 has some issues, including introducing an inconsistency between EO and BO coding, having a larger amount of changes than expected. – This is resolved in the suggested text of I0602.)
Decision: (for SAO cleanup) [Remove decision label here?]
JCTVC-I0602 BoG report on integrated text of SAO adoptions on top of JCTVC-I0030 [Y.-W. Huang, E. Alshina] New text provided in JCTVC-I602 This contribution integrates related texts of SAO adoptions in the 9th JCT-VC meeting. The integration includes JCTVC-I0168, JCTVC-I0066, JCTVC-I0172, removal of SAO APS mode, fix of ticket #504, ticket #517, fix of ticket #518, and JCTVC-I0856. Based on the text provided in JCTVC-I0030 [1], the SAO related texts are incrementally modified as follows;
[Are the above square-bracketed numbers useful/appropriate in the meeting notes?] In principle, Adopt SAO text changes as suggested in JCTVC-I0602. One issue about the position of sign coding (before or after magnitude, sign coded in bypass mode) is still open and experiments are running. The current version of I0602 encodes sign first, which seems to be inconsistent as it useless when the magnitude is zero. Decision: After the experiments were finished, this change is adopted. Additional note: Some people understood that an earlier version of I0168 claimed to use context coding of the sign, but actually the software implementation used bypass mode, and this was clarified in discussions and in a later revision. It is reported verbally that the performance is not substantially changed if context mode would have been used. No continuation of CE on SAO, perhaps AHG for further study on stabilization and alignment between text and software (further “simplifications” need sufficient evidence about benefit). Decision (SW): Use SAO settings as in CE1 (disable some of the encoder tricks, see H1101) as default configuration. Two documents were presented in track A in conjunction with BoG report: I0507 (presented before in BoG but presentation supported by 1 non-proponent) In the HM6.0 process, the coding of the SAO merge left and merge up is performed for each color component. This contribution proposes to use the merge left flag to signal whether all color components are merge left, and to use the merge up flag to signal whether all color components are merge up. Otherwise, SAO offset parameters are signaledsignalled for each color component. One concern expressed how this can be combined with RDO. RDO must be performed separate for the three components which makes it difficult to come to the optimum point. Advantage is that no conditional parsing is necessary. The proponent claims that the performance is equivalent to I0199, but for LCU=64 it is 0.3% worse than JCTVC-I0199. Late contribution. No action.
This document proposes a modification to SAO type coding. Only two types are considered: Edge Offset (EO), one Band Offset (BO). Therefore, a 1-bit flag is used to signal the SAO type. The SAO on/off switch is also coded using a 1-bit flag with a separate context. The rest of the side information is coded using Fixed Length (FL) codes and bypass mode in CABAC. No draft standard text proposal had been available until 04-28 23:55; nobody has studied it. The cross-checker supported the proposal, but had not been given the draft text before and could not confirm its quality. The draft text presented in track A contained a substantial amount of changes. No action, as this is a new proposal and stability of the standard has highest priority. JCTVC-I0580 BoG report on review of non-CE ALF proposals [A. Norkin] The proposals were sorted into categories several categories.
Recommendation of BoG:
[Clarify above w.r.t. which are BoG rec. vs. JCT-VC actions (only the second bullet is a JCT-VC action).] JCTVC-I0585 Report on subjective viewing of ALF [T. Yamakage (Toshiba)] (insert abstract) Report of ALF Viewing verbally given Monday evening About Tile boundary (I0167): According to visual testing, no visual difference or proposal very slightly worse (BB drill QP37) Conclusion: No action, Decision: Do virtual boundary processing also at picture boundary consistent with any other boundary About general ALF quality: Out of 40 test cases, 6 showed difference as per confidence intervals above zero, in all except for one cases, ALF was judged better. In plenary discussion, it was mentioned that two of the test points where ALF is better are with Riverbed, where it is known that this sequence has some special characteristics (see also discussion under deblocking filter). The version of ALF tested here is the implementation according to the simplified description I0603. It was run with the "low delay" constraint for derivation of filter coefficients, i.e. the coefficients are optimized for the previous encoded picture.
(include abstract). Grouping of proposals: I0075, I0351 (MV restriction) – preferred solution is one where the decoding process implicitly imposes encoder constraints, which is fulfilled by I0351 case 1.1 I0107, I0297, I0425 (PU restriction) – solution is preferred where the change is applied after MV reconstruction. I0297 case 1 selected, which achieves this without syntax change. I0351 has less BD rate loss, but more text changes. I0297 changes the merge part and MV reconstruction, I0351 the MV prediction and reconstruction. Worst loss is for 1 sequence is 0.8/0.9%. Wed. afternoon Broadcom (TH) reports about a further analysis which unveiled that I0297 achieves an effective worst case memory bandwidth reduction by 50%, whereas I0351 stays around 35%; bi-prediction still requires more memory accesses even if reduced to full-pel in one direction. Decision: Adopt JCTVC-I0297 case 1 This feature is always on, no flag at SPS. The encoder can indicate bipred, but the decoder reacts, at least to some extent, as if it did not. It was onfirmed by B. Bross that the syntax change is OK, however it is in principle not necessary to set the MV to zero (editorial).
Notes for each topic reviewed in this BoG were integrated into the sections in this report on each of the reviewed contribution documents. JCTVC-I0581 DRAFT: BoG report on high level parallelization [Andrew Segall] Notes for each topic reviewed in this BoG were integrated into the sections in this report on each of the reviewed contribution documents. [Rearrange notes of documents to more closely follow the order in the BoG report.]
Recommendation of BoG:
Decision: Adopt using the average of QP as suggested by BoG. Not adopt the bitstream constraints, as imposing such constraints would not have any advantage for implementing normative decoder behaviour, and not observing these constraints would not cause a decoder crash. JCTVC-I0587 BoG report on SAO: Results of subjective and objective tests [Y.-W. Huang] From previous discussion:
Tests were conducted in a way that three sequences were shown, and then subjects were asked the “better-equal-worse” question for A versus B, and A versus C. Such a test would only make sense when A would be a common anchor. Test needs to be re-done. For the second round of testing, JCTVC-I0137 was withdrawn, as the proponents recognized that JCTV-I0247 is similar but simpler. I0073: No loss in visual quality (confidence intervals overlapping, very slight improvement in one case Johnny QP 37) I0162: No loss in visual quality (confidence intervals overlapping) (I0073 and I0162 presented again in track A) I0073:
This would make the intended unification of edge offset and band offset coding different again (see under I0576) I0162:
Some concern is raised whether this would limit the degrees of freedom. This is not in conflict with the unification of edge and band offset coding that was adopted. Note: Production of a cleaned and clarified SAO description with the unifications of edge and band offset coding has highest priority (see under I0573). Therefore, the suggestion in 0073 appears not to simplify. The total memory needed for SAO line buffers in the current design is not critical. Reductions should not reduce the freedom in the ways of operating SAO (e.g. reducing number of offset values). I0247: No loss in visual quality – no action. I0184: No loss in visual quality (confidence intervals overlapping, perhaps one case Basketball QP32 very very slightly better) I0199: No loss in visual quality (confidence intervals overlapping, perhaps one case Party Scene QP32 very very slightly better) I0199 moves the syntax element for SAO on/off before the merge flag, introducing one more context and one more dependency. When neighbor is on/off, SAO_merge is not parsed. This gives 0.4% gain. I0184 is non-normative. This gives 0.2% gain. Combination of both gives 0.4% gain. The additional gain of I0199 versus the additional parsing dependencies does not justify its consideration.
Further study whether the merge flag can be skipped (current results indicate that this comes with some performance drop). JCTVC-I0593 BoG Report: Extended Adaptability Range on Deblocking Filter [T. Suzuki] There were proposals to strengthen deblocking filter. JCTVC-I0071 proposes to extend β, and JCTVC-I0258 proposes to modify tC table by setting higher values in the high QP region (outside the common conditions test range). This BoG was established to explore suggest a solution for extended adaptability range of deblocking, candidates being I0071 #1 (without additional syntax) and I0258. Subjective tests were performed with 4 sequences (sequences outside common test set, where it was expected that deblocking would become more visible).
From JCTVC-I0071: method 1 beta_k_log2=2 or 3 From JCTVC-I0258: Replace tc table by proposed tc in I0258
HM with default offset versus combination of 0071 and 0258, both cases beta_k_log2 Riverbed: QP32 ControlledBurn: QP27, QP37 WestWindEasy: QP32 ChinaSpeed: QP32 No visual improvement was found (some blocking artefact in anchor, but not significant, some more blurring in the alternative solution. Differences became more visible when the contrast of the display was increased (which may not be an appropriate way to compare).
One expert reports that most visible difference (less blocking, more blurring) appeared between the two HM6 cases, whereas the other two cases of combined proposals introduced some more blurring. BoG recommends
Yüklə 1,12 Mb. Dostları ilə paylaş:
|