International organisation for standardisation organisation internationale de normalisation
Yüklə 8,24 Mb.
|
5.8.2Sample adaptive offset5.8.2.1.1.1.1.1.1JCTVC-H0067 Non-CE8.a.2: SAO with LCU-based syntax [A. Fuldseth, G. Bøntegaard (Cisco)] This scheme was proposed as an alternative approach to that in H0273 (also proposed as F056 and G831. Asserted simplifications included the following:
The syntax elements are proposed to be interleaved at the LCU level (which would make it harder to perform whole-picture optimization). The proponent indicated a code size reduction by 3000 to 4000 lines of code (but it was remarked that the resulting software does not support the full required functionality – e.g. interactions with tiles, variable-size slices, non-CTC LCU sizes). Only a single CABAC context was used (as proposed) – there could be gain from using multiple contexts (e.g., different contexts for different syntax elements). Coding efficiency impact: AI/RA/LB/LP/RA_10 = -0.1/0.3/0.8/0.7/1.1% for CTC. It was commented that Class F has significant losses. It was commented that the edge offset limitation harms the 10-bit case. It was commented that if this scheme is compared with whole-picture optimization of H0067, significant loss occurs. It was suggested that whether the data is interleaved or placed in the slice header could be controlled by a flag. Doing that, with a single decoder signal processing operation that is invariant to the location of the sent data, seems like a good idea. It was suggested to test H0273 with a such a flag during this meeting and discuss the subject further. Further study of other aspects was also suggested. That testing was reportedly performed (using 3 added contexts), and was reportedly cross-checked by TI. It was commented that using the bypass mode in CABAC might be preferable. Without interleaving into the SD, the effect on coding efficiency was 0.1/0.1/0.3/0.3% degradation for AI/RA/LB/LP. With interleaving into the SD, the reported effect was 0.0/0.2/0.4/0.6% on luma, with greater losses for chroma. This version still used 16 offsets, and the adopted reduction of the number of offsets to 4 should mitigate the losses reported with interleaving. The proposal limits the maximum number of different offsets per picture to 1024. The HM uses a quadtree with a maximum depth of 4 that does not reach below the LCU level. This would probably be more efficient when the LCU size is smaller. The use of bypass mode should be studied. Decision: Adopt this modified form of H0273. When not interleaved, the control data should be sent in the APS (not using a quadtree), and the APS must be static for the entire picture. It was remarked that the APS cannot be parsed without the SH or SPS. Decision: Put a loop quantity indicator in the APS as necessary to enable parsing it without the SH or SPS. It was noted that the text impact for the H0273 and H0274 related adoptions is relatively high. The contributors worked with the chief editor during the meeting to allay these concerns for H0273. Concerns remained expressed for H0274; however, the contributors indicated their willingness to substantially assist with editorial improvement of the text for H0274. In was noted that the availability of suitable text is a fundamental requirement in our work (see also section 7.1 in this regard). Further study of additional simplifications (e.g. of H0067), quadtree syntax, and other aspects in CE or AHG was requested. 5.8.2.1.1.1.1.1.2JCTVC-H0716 Crosscheck of JCTVC-H0067 on "Non-CE8.a.2: SAO with LCU-based syntax" [C. Auyeung (Sony)] [late 02-04]
This contribution was discussed as recorded above in relation to CE8 related contributions. 5.8.2.1.1.1.1.1.6JCTVC-H0183 Non-CE8: Cross-check report for sample adaptive offset with constrained ranges of edge offset (JCTVC-H0114) [T.Chujoh (Toshiba)] [late]
In this contribution, performance test results were provided for enabling Sample Adaptive Off-Set (SAO) relative to the LC test configuration. On top of HM5.0, SAO reportedly provides 1.0% (AI-LC), 2.1% (RA-LC), 3.6% (LD-LC), 8.4% (LP-LC) average Luma BD bit rate gain, and some higher Chroma BD bit rate gain. The motion compensation (MC) part complexity measurement reportedly shows 2% reduction of multiplications and 1% average memory bandwidth reduction when SAO is enabled. It was asserted that the denoising effect of SAO caused the encoder to choose the "more difficult" motion compensation fractional positions less often, thus achieving an asserted reduction in overall average complexity of the motion compensation process (as measured by multiplies, adds, and memory accesses). The cross-checker confirmed that the measurements were accurate and that the statistics of the motion vectors had changed. The cross-checker also suggested that SAO may be useful for low-complexity ME encoders that do not fully search for fractional positions. The contributor indicated that similar effects are observable with ALF (although ALF has more complexity). The contributor proposed to include SAO in LC configurations. 5.8.2.1.1.1.1.1.8JCTVC-H0579 Cross-verification of sample adaptive offset in low complexity test (JCTVC-H0115) [K. Chono, H. Aoki (NEC)] [late]
In this contribution, some visual artefacts associated with SAO in HM5.0 were reported, and an approach to remove the artefacts was proposed. The artefact was reportedly observed at SAO partition boundaries and category boundaries. In the proposal, a "weak" offset value is proposed to be added to reconstructed samples along the boundaries. It was reported that the artefact was not observed when this proposal is applied on top of HM5.0. Experimental results reportedly showed that the proposed changes can be introduced on top of HM5.0 with 0.1%-0.3% average BD bit rate loss for the common test conditions. Two methods were proposed for the artefact removal:
The proposed approaches might seem to increase complexity. The cross-checker said that the category boundary approach would not add much complexity, but he did not know which might be enough or whether both might be needed. It was commented that the artefact is associated with band offsets, that a modification of the band offsets had been adopted that may mitigate the problem, and that an encoder constraint to avoid large differences at boundaries may be sufficient to deal with the problem. We would definitely want to conduct further study to determine whether there is still such a problem, and if there is a non-normative way to fix it (without degrading visual quality), we would want such a fix in our reference software encoder. Further study was certainly encouraged. AHG or CE work on the study of visual quality of in-loop filtering was planned. 5.8.2.1.1.1.1.1.10JCTVC-H0439 Cross-verification of Non-CE8: Method of removing boundary artefacts for SAO (JCTVC-H0201) [W.-S. Kim, D.-K. Kwon (TI)] [late]
This contribution was discussed as recorded above in relation to CE8-related contributions. 5.8.2.1.1.1.1.1.12JCTVC-H0338 Non-CE8: Crosscheck for Sharp and TI BO in JCTVC-H0438 [C.-M. Fu, Y.-W. Huang (MediaTek)] [late]
In the HM5.0 process, an offset is added to each pixel in attempt to reduce the distortion of the reconstructed pixel relative to the original pixel. In this contribution, the operation in which the offset is applied to the pixel is modified, and the offsets are quantized. See also the related contribution H0557. It was commented that the analysis seemed correct and would resolve any potential "salt and pepper noise" issue, but would increase complexity. The contributor indicated that, with the removal of the sign bit, the complexity of the clipping might be less. Further study was highly encouraged. 5.8.2.1.1.1.1.1.14JCTVC-H0553 Non-CE8: Harmonization of LCU-based SAO [K. Minoo, Y. Yu, D. Baylon, L. Wang (Motorola Mobility)] This contribution was discussed as recorded above in relation to CE8-related contributions. No simulation results were in the contribution. A revision was later uploaded to include simulation results. The decoder would determine a min/max range over the LCU, avoiding the need for signalling. It was remarked that this seemed promising. Further study was encouraged in a CE or AHG. 5.8.2.1.1.1.1.1.15JCTVC-H0708 Non-CE8: Crosscheck for Motorola Mobility harmonization of LCU-based SAO in JCTVC-H0553 [C.-M. Fu, Y.-W. Huang (MediaTek)] [late 02-05] 5.8.2.1.1.1.1.1.16JCTVC-H0406 Flexible Band Offset Mode in SAO [E. Maani, O. Nakagami (Sony)] This contribution was discussed as recorded above in relation to CE8-related contributions. No working draft text was in the contribution. 5.8.2.1.1.1.1.1.17JCTVC-H0617 Cross-verification of Flexible Band Offset Mode in SAO (JCTVC-H0406) [W.-S. Kim, M. Zhou (TI)] [late]
Yüklə 8,24 Mb. Dostları ilə paylaş:
|