Organisation internationale de normalisation
Block Structures and Partitioning
Yüklə 7,54 Mb.
|
17.6Block Structures and Partitioning17.6.1.1.1.1.1.1.1JCTVC-D087 Encoding complexity reduction by removal of NxN partition type [Jungsun Kim, Byeongmoon Jeon] This contribution proposes removal of the NxN PU partition type for both inter and intra predicted CUs to reduce encoding complexity. In the current HM, a CU may split into several PU partitions. Two types of splitting and four types of splitting are used for intra and inter CUs, respectively. In the proposed methods, the NxN PU partitioning is prevented except when the CU has the maximum depth. The proposed scheme reportedly shows approximately 29% encoding time reduction with 0.2% loss in BD BR under the condition of intra high efficiency and 18% encoding time reduction with 0.3% loss in BD BR under the condition of random access high efficiency. Corresponding losses for LC were 0.9% intra, 0.8% inter. 17.6.1.1.1.1.1.1.2JCTVC-D432 Remove Partition Size NxN [S. Liu, Y.-W. Huang, S. Lei] (late registration Friday 21st after start of meeting, uploaded Friday 21st, second day of meeting) This contribution proposes to remove the PU partition size NxN for both inter and intra CUs, for the purpose of removing encoding redundancies. Both inter and intra CU syntaxes are modified accordingly. Results reportedly show that with this modification the encoding time can be reduced by 20% and 30% for inter and intra, respectively, at 0.2% BD BR increase for HE. Results were similar to JCTVC-D087 (slightly less loss reported) It was noted that the loss is higher in LC case. The question was asked whether it would be better to set RQT depth = 2 to keep some of the performance. For LC, LCEC was replaced by CABAC in LC, and it was found that the loss came to a similar range as for HE, so it was concluded that this is a problem of LCEC. It was asked whether the same goal could be achieved by just modifying the encoder, while keeping the syntax unchanged. Two experts expressed the opinion that, for the intra coding case, removing NxN mode for CU’s larger than the minimum size may restrict some functionality (e.g. related to copying chroma modes from luma modes). It was reported that by handling this as an encoder-only issue (i.e. retaining the syntax), the BR increase would be around additional 0.1% for HE and 0.5% LC case Decision: Remove NxN for CU’s larger than minimum size also for intra, for both HE and LC configurations. The problem with LC mode should be further investigated. Additionally, a request was made to remove the NxN partitioning for inter at the minimum CU level. Arguments are made against this ‒ that it may be disadvantageous for high bit rates. 17.6.1.1.1.1.1.1.3JCTVC-D060 Evaluation results on Residual Quad Tree (RQT) [Minhua Zhou, Ali Tabatabai] This document reports evaluation results of the residual quadtree (RQT) adopted in the first version of the HEVC test model (HM1). When compared to two-level method used in TMuC-0.7, the RQT reportedly provides marginal gain in the luma component of high efficiency configurations, but reportedly leads to loss in other configurations, especially in the chroma component parts in which the significant loss was reported. The RQT performance with respect to encoder complexity was also discussed in this contribution. It was asserted that the current RQT design needs further improvement in regard to the complexity and coding efficiency performance as tested and recommended to set up an ad-hoc group for the development of improved TU tree representation methods which would simplify both the encoder and decoder design.
Further study in an AHG was suggested. 17.6.1.1.1.1.1.1.4JCTVC-D250 Efficient transform unit representation [Krit Panusopone, Xue Fang, Limin Wang] This document proposes a modified transform unit representation. The proposed approach modifies the RQT based-TU partitioning scheme in the TMuC. The proposed syntax change offers a particular trade-off between coding efficiency and additional complexity. It is proposed that maximum TU size be linked to PU size, instead of CU size and RQT be completely removed, i.e. for the TU size to always be the same as the PU size (or, for non-rectangle PUs, to have 2 TUs in the PU). It was only suggested to apply this for inter, as the loss for intra would be too high (see TE12 from the third meeting). Losses are reported as 0.9/0.9% BD increase for RA, and 1.0/1.8% for LD (HE/LC). Encoding time was reportedly reduced to 75/90%.
17.6.1.1.1.1.1.1.5JCTVC-D433 Cross-check for Motorola Mobility’s Proposal JCTVC-D250 on Efficient Transform Unit Representation by RIM [Jinwen Zan (RIM), Dake he (RIM)] (late registration Saturday 22nd after start of meeting, uploaded Saturday 22nd, third day of meeting) 17.6.1.1.1.1.1.1.6JCTVC-D142 A unified design of RQT cbf coding in LCEC [B. Li (USTC), J. Xu (Microsoft), F. Wu (Microsoft), H. Li (USTC)] In the TMuC 0.9, RQT is used to represent the residue. RQT is more flexible than a fixed number of transform levels. When representing the residue in RQT, besides the residue coefficients, the quad-tree structure and the cbf related to the quad-tree structure need to be signaled to the decoder. When LCEC is used, a combined coding of the cbf is used in the current TMuC. But in the current design, when max transform depth > 2, disabling the combined coding will reportedly get better results. This contribution presents a design for both max RQT depth = 2 and depth > 2, which has different performance for both cases. Compared with the reference configurations, it was asserted that the proposed method will improve the coding efficiency with no appreciable impact on complexity at both encoder and decoder side. The split flag is coded as usual. In the case of depth 2, there is negligible change. In the case of depth 3, combined coding was disabled, as it reportedly works better. Compared to that, the reported gain is larger. In discussion, the general question was asked wheter the restriction of 2 levels would be retained. 17.6.1.1.1.1.1.1.7JCTVC-D375 LCEC coded block flag coding under residual quadtree [X. Wang, W.-J. Chien, M. Karczewicz (Qualcomm)] This proposal presents a coding scheme for coded block flag under the residual quadtree structure. Based on the proposal, the current joint coding of Y, U and V coded block flag at CU root level is extended in two ways: 1) it further includes transform split flag if available; 2) The joint coding is not only performed at the CU root level, but also extended to a higher transform depth level so that the benefit of such joint coding can be maximized. Simulation results reportedly show that such extension of the joint coding can improve coding efficiency, especially when there are multiple levels of transform tree decomposition. The current implementation of CBP coding under LCEC reportedly works well with a maximum TU depth level 1 and 2. When the maximum depth level is 3, the performance is reported not to be so good. Some investigation into the issue was reportedly done after release of TMuC 0.9 and it was reported that it is not because of the CBP signaling mechanism, but the RD not quite matching to CBP signaling. With changes in the RD function, the coding performance of the current CBP signaling under LCEC reportedly works well too for a maximum depth level of 3 or higher.
The following approach was suggested for more efficient coding of CBP:
Decision: Adopt the bug fix (encoder only issue) – this would resolve the current warning message. Further investigate in a CE (JCTVC-D142, JCTVC-D375) further improvements of 3-level configuration in LC, 17.6.1.1.1.1.1.1.8JCTVC-D143 Results for redundancy reduction in PU-based merging [B. Li (USTC), J. Xu (Microsoft), F. Wu (Microsoft), H. Li (USTC)] In JCTVC-C277, a method was proposed to reduce the motion vector candidates in CU/PU-based merging. This document studies the performance of such a method in PU-based merging with TMuC0.9. When CU is SIZE_2NxN, SIZE_2NxnU or SIZE_2NxnD, the motion of the PU above a location cannot be a candidate for the PU below it. When CU is SIZE_Nx2N, SIZE_nLx2N or SIZE_nRx2N, the motion of the PU on the left cannot be a candidate of the PU to its right. This change had been adopted according to JCTVC-D441. 17.6.1.1.1.1.1.1.9JCTVC-D165 CU-level Directional Merge Mode [Jicheng An, Xun Guo, Yu-Wen Huang, Shawmin Lei] In this contribution, a directional merge method is proposed to improve the CU-level merge mode in HM1.0. The syntaxes of two merge methods are the same. The contribution changes their semantics to use the motion information from more neighboring blocks to increase the chance of sharing motion parameters from neighboring blocks. Compared to the anchor of HM1.0, average bit rate reductions of 0.2%, 0.3%, 0.4 and 0.5% were reported for random access HE, random access LC, low delay HE and low delay LC, respectively. If compared to HM1.0 with merge mode off, the proposed method can reportedly achieve bit rate reductions of 0.8%, 0.2%, 1.2 and 1.6% for random access HE, random access LC, low delay HE, and low delay LC, respectively. There is reportedly essentially no increase in encoding time and decoding time for the proposed method. Motion vectors from a row of blocks above or from a block of columns to the left are merged – with a reported increase of complexity being very small – several experts supported to further investigate this. It was asked how it works with the new design (PU based merging) Further study in CE9 was suggested. 17.6.1.1.1.1.1.1.10JCTVC-D316 Crosscheck of MediaTek's Proposal JCTVC-D165 on CU-level Directional Merge Mode [B. Bross (Fraunhofer HHI)] Results were confirmed and further study suggested 17.6.1.1.1.1.1.1.11JCTVC-D249 Super large coding tree block [Krit Panusopone, Xue Fang, Limin Wang] This document describes the concept of "super largest coding unit" (SLCU) to allow representation of largest coding unit (LCU) in a more efficient way. It was asserted that this reduces the overall quadtree overhead for high resolution video as well as the complexity for low resolution video. An SLCU is proposed, consisting of 2x2 regular LCUs. The processing order is different from normal LCU order. The same split flat syntax is used for LCU and below in the coding structure. A split flat syntax is proposed for the SLCU level:
LCU size 32 was used for classes C and D. Results in 0.3%/0.4% BR increase, mostly from classes C and D, with encoder speedup by roughly 10%. It was asked whether this speedup is mostly at low or high rates. Indeed the saving is reported to be mostly in classes C and D; when setting max CU size = 32, the saving would be marginal. The benefit of the SLCU structure and associated syntax elements does not appear obvious. Yüklə 7,54 Mb. Dostları ilə paylaş:
|