Non-CE Technical Contributions (164) Range extensions (87) General (1)
14.1.97.1.1.1.1.1.170JCTVC-P0288 BoG report on Range Extensions [C. Rosewarne, K. Sharman]
Discussed in JCT-VC plenary Sunday 01-12 a.m. (JRO & GJS).
Further discussed in JCT-VC Wed 01-15 (GJS).
A break out group was held between Thursday the 9th and Monday the 13th of January 2014 to review input documents relating to the Range Extensions. Specifically, topics on RCE1 on entropy coder throughput, RCE2 on Rice parameter initialization and RCE3 on intra block refinement and other technical contributions were discussed.
In addition, contributions indicated that RExt is giving significant compression benefits over AVC High 4:4:4, with up to 8:1 compression improvement reported for screen content (see P0200 and P0213), and around 30% for AI, and 40% to 50% for inter (based on PSNR).
This report conveys a number of recommendations from the BoG. These recommendations are summarized in section 1. Throughout the report, the expression "Recommendation:" was used to express the consensus in the group when a positive action was recommended by the AHG.
(Note: RCE4 was not addressed in the BoG.)
First pass on contributions had been completed.
The current status was reported by the BoG and discussed by JCT-VC:
RCE1:
-
(updated 01-15 and 01-16) If there is a 16-bit profile with a "super tier", using "Option B3a" (setting CABAC to raw bin coding if any coeff_abs_level_remaining codeword is present). An SPS level flag would indicate this usage, and would be required to be 1 in such "super tier" bitstreams – and this requires consideration of whether "super tier" decoders would also be required to decode lower-tier bitstreams. It was remarked that the lower-tier decoding capability requirement might not provide the decoder with the intended complexity reduction, and that this might mean that some profile might be needed that would not be required to decode bitstreams that don't use this feature (e.g., an all-intra profile that has a non-onion-shell relationship with other all-intra profiles). Decision: Agreed (after parent-level review regarding profiling about the need for such a profile, which left the question open regarding whether the decoder would be required to decode lower-level bitstreams). Add control flag and low-level specification, prohibit usage in non-16 bit profiles, identify as TBD in 16 bit profiles, identify 16 bit 4:4:4 as TBD.
RCE2:
-
(updated 01-15 and 01-16) Rice parameter initialization – high bit-depth lossless coding had not originally been considered (there was interaction of tested approaches with new AHG18 lossless test results). After consideration of this issue, it was suggested to focus on the A1 scheme of P0199 with unlimited Rice parameter. Decision: Adopted as described.
RCE3:
-
Padding approaches for intra block copy are not necessary.
-
Initialization of the intra block copy vector predictor is left unchanged. In follow-up discussion about initialization in the JCT-VC plenary Sunday a.m., a contributing participant requested re-consideration of the initialization issue (which currently initializes to zero, which cannot be a valid displacement). As a result of this follow-up discussion, an initialization discussed in JCTVC-P0111 (initialization to displacement (−W,0)) was adopted. Decision: Initialize to (−W, 0).
(Test results for this initialization value are provided in P0217-v2.)
-
The allowed range of vectors for intra block copy was not recommended to be decreased.
-
Clarification was requested regarding all discrepancies between the intra block copy software and the text. In the JCT-VC review, this was agreed as follows:
-
The intended behaviour of the BV range for different CTU sizes (i.e., always 64 wide versus the current CTB width). Decision: Always 64 wide for luma.
-
Reconstruction process (TU or CU based). The current text already has a restriction that would allow the decoder to use CU-based prediction.
-
(updated 01-15) Adoption of either of: “NxN only” or “NxN/2NxN/Nx2N” partitioning for intra block copy (with PU-level overlap prohibited). Text for “NxN/2NxN/Nx2N” was provided in an update to P0180. Between these two, the “NxN/2NxN/Nx2N” variation was preferred in plenary discussion. No concerns were raised regarding adoption of this.
JCT-VC discussion chaired by C. Rosewarne: Decision: Adopted.
-
The earlier discussion was opened for further comments and no new concerns were raised.
-
(updated 01-15) To establish an RCE if NxN intra block is adopted (i.e. either of the above options) to examine block vector coding and binarization, to include P0149. To be further discussed to decide whether to define an RCE.
Non-CE:
-
The BoG requested review of the remaining two contributions under “Source video test material”.
-
The BoG recommended to discuss establishing a CE to study worst case code lengths for coeff_abs_level_remaining for high bit-depths and complexity aspects. [P0061]
-
Decision: Adopt P0066 on bit depth correction for CCP.
-
The BoG recommended plenary-level discussion for SEI for FPA. See section 7.1.
-
Add lossless testing to the AHG18 test conditions and/or establish a CE relating to this.
-
(updated 01-16) Decision: Adopt change to SAO for high bit depths (P0222) to send shift values for luma and chroma for the offset value in PPS (no change to classification, with shift value in the range of 0 to Max{BitDepth – 10, 0}).
-
To discuss establishing a CE/AHG for motion vector precision
-
To discuss establishing a CE/AHG to examine encoder configuration for CCP, regarding two methods (predict Cr from Cb, and use of a programmable LUT)
-
Decision: Adopt an additional context for CCP alpha signalling [P0154].
-
Further study is recommended on encoder RQT searching following previous changes for CCP.
-
See other notes regarding P0044.
-
(updated 01-16) The group considered moving the rotation and single context enable from the SPS to the PPS conditional on trans-quant bypass or transform skip being enabled [P0166], but thought that having these aspects switching on a picture-by-picture basis did not seem reasonable, so no action was taken on this.
Further discussion in closing plenary (JRO):
-
Integer MVs for inter prediction gives benefit for screen that was likely rendered with integer precision. Requires changing the interpretation of decoded motion vector (conditional shift), and conditional rounding of predictor. No change of parsing. This appears as a slight increase of worst case complexity, whereas for content types where it is used the average complexity would decrease. Agreed to start a CE. Part of the reported gain could likely be achieved by encoder decision, this should also be investigated in the CE.
-
TU or PU overlap IBC - More study is required on the complexity impact. No CE.
14.1.97.1.1.1.1.1.171JCTVC-P0166 Some Syntax Modifications for HEVC Range Extension [Y. Yu, L. Wang Arris)]
See BoG report P0288 and related notes.
Dostları ilə paylaş: |