Organisation internationale de normalisation
Yüklə 7,54 Mb.
|
10.2MDDT simplification10.2.1.1.1.1.1.1.1JCTVC-D107 (m18858) CE7: Experimental results for one-dimensional directional unified transform (JCTVC-C080) [Akiyuki Tanizawa, Jun Yamaguchi, Taichiro Shiodera, Takeshi Chujoh, Tomoo Yamakage (Toshiba)] This contribution presented experimental results of the use of a 1 Dimensional Directional Unified Transform (1DDUT) for Core Experiment 7 (Alternative transforms in HEVC). 1DDUT is a spatial transform scheme for intra coding. 1DDUT has only two 1-D transform matrices, which are one similar to a Discrete Cosine Transform (DCT) and the directional transform based on Karhunen-Loeve Transform (KLT). 1DDUT was proposed in the contribution of JCTVC-B042 at the JCT-VC Geneva meeting and JCTVC-C080 at the JCT-VC Guangzhou meeting. In this document, experimental results in TMuC software version 0.9 on both high efficiency conditions and low complexity conditions based on the common test conditions for both I slice only coding structure and the random access coding structure defined by CE7 were reported. For I slice only coding structure, the BD-Bit rate gain compared with the anchor was reported as 2.2% on average, and for the random access coding structure, the BD-Bit rate gain was reported as 0.8% on average. Experimental results reportedly indicated that 1DDUT can improve the coding efficiency for intra coding with minimum complexity increase.
10.2.1.1.1.1.1.1.2JCTVC-D032 (m18775) CE 7: Cross-Check for Toshiba’s proposal on 1-D Directional Unified Transform by Samsung [A. Saxena, F. C. Fernandes (Samsung)] 10.2.1.1.1.1.1.1.3JCTVC-D264 (m19028) CE7: A crosscheck of Toshiba's proposal (JCTVC-D107) by Qualcomm [R. Joshi, P. Chen] 10.2.1.1.1.1.1.1.4JCTVC-D046 (m18793) CE7: Mode-Dependent Transforms for Block-based Intra Coding from Institute for Infocomm Research [C. Yeo, Y. H. Tan, Z. Li, S. Rahardja (I2R)] To simplify the operation of Mode Dependent Directional Transform (MDDT), a Mode-Dependent Fast Separable KLT for Block-based Intra Coding was previously introduced that requires two transform matrices: one like a DCT and one similar to a DST (Odd Type-3 DST). The 4x4 DST reportedly also has a structure that can be exploited to reduce the operation count of the transform operation. This contribution provided test results for CE7. Experimental results reportedly showed that the proposed technique matches the performance of using trained KLTs even though the approach requires no training, and has lower operation count and storage costs.
10.2.1.1.1.1.1.1.5JCTVC-D104 (m18855) CE7: Cross check report of I2R's proposal (JCTVC-D046) from Toshiba [A. Tanizawa, J. Yamaguchi (Toshiba)] 10.2.1.1.1.1.1.1.6JCTVC-D306 (m19072) CE7: Crosscheck of I2R’s proposal by Huawei [H. Yang, J. Zhou] 10.2.1.1.1.1.1.1.7JCTVC-D353 (m19121) CE7: Cross-verification of I2R’s experimental results of low complexity MDDT by Samsung [E. Alshina, W.-J. Han, A. Saxena (Samsung)] Confirmed results; mentioned that for 8x8, the transform is implemented by matrix multiplication. 10.2.1.1.1.1.1.1.8JCTVC-D033 (m18776) CE7: Mode-dependent DCT/DST for intra prediction in video coding [Ankur Saxena, Felix C. Fernandes] It was previously reported that following intra prediction, the optimal transform is a Discrete Sine Transform (DST) with performance close to KLT, along the direction of prediction. In this contribution, the proponent used a transform like a Discrete Cosine Transform (DCT) or DST (separably along the vertical or horizontal directions) based on the intra prediction direction. The proposed DCT/DST transform scheme does not require any additional signaling information or Rate-Distortion search during encoding, and works in a single-pass. No training was required to derive the transform and the scheme requires the storage of only one DST matrix in addition to the conventional DCT-like transform at each block size. The conventional quantization tables for TMuC 0.9 are retained and no changes had been made to the scanning order. Experimental results were provided with TMuC 0.9 as anchor for the test conditions as stipulated in Core Experiment 7, and average BD Rate gains of 1.3%, 2.1%, 0.6% and 0.7% were respectively reported for Intra High Efficiency, Intra Low Complexity, Random Access High Efficiency and Random Access Low Complexity settings. DST and DCT (once with kernel from HM block sizes 4x4 and 8x8, once with matrix multiplication block sizes up to 32x32) 6 quantization and inverse quantization matrices each were used. No fast DST was used, and usage would require 3 additional quantization matrices. For the case of only 4x4 and 8x8 transforms, BR reduction was reportedly 1.2%/1.9% for intra HE/LC cases, with encoder runtime increase to 104%/97%, decoder 102%/99%. Reported by cross-checkers, the runtimes are higher up to 110% also for LC case, and the proponent agrees that the numbers reported may not be reliable. 10.2.1.1.1.1.1.1.9JCTVC-D031 (m18774) CE7: Cross-check for Samsung's Proposal on Jointly Optimal Intra Prediction and Adaptive Primary Transform by BBC [Ying Weng, Thomas Davies (BBC)] 10.2.1.1.1.1.1.1.10JCTVC-D088 (m18838) CE7: Cross-check for Samsung's Proposal on Mode-dependent DCT/DST for intra prediction in video coding by NHK [Yasuko Sugito, Atsuro Ichigaya] 10.2.1.1.1.1.1.1.11JCTVC-D105 (m18856) CE7: Cross check report of Samsung's proposal on mode-dependent DCT/DST (JCTVC-D033) from Toshiba [Akiyuki Tanizawa, Jun Yamaguchi (Toshiba)] 10.2.1.1.1.1.1.1.12JCTVC-D399 (m19178) CE7: Mode dependent intra residual coding analysis [P. Chen, R. Joshi, Y. Zheng, M. Coban, M. Karczewicz] In this contribution, several different methods of mode dependent intra residual coding are compared. They are mode dependent fix scan, adaptive switch scan and adaptive switch KLT+scan. The trade off between coding efficiency and complexity was discussed in the contribution.
10.2.1.1.1.1.1.1.13JCTVC-D078 (m18828) CE7: Cross-check of Qualcomm’s proposal by Institute for Infocomm Research [C. Yeo, Y. H. Tan, Z. Li (I2R)] 10.2.1.1.1.1.1.1.14JCTVC-D307 (m19073) CE7: Cross-checking of MDDT results from Qualcomm [J. Song, H. Yang] (initial version rejected as a placeholder upload) 10.2.1.1.1.1.1.1.15JCTVC-D354 (m19122) CE7: Cross-verification of Qualcomm’s MDDT [E. Alshina, V. Seregin, W.-J. Han (Samsung)] 10.2.1.1.1.1.1.1.16JCTVC-D392 (m19171) CE7: Mode dependent intra residual coding - A joint proposal based on several proposals from CE7 [R. Joshi, P. Chen, M. Karczewicz, A. Tanizawa, J. Yamaguchi, C. Yeo, Y. Tan, H. Yang, H. Yu] This proposal suggests, for intra predicted blocks, for the prediction residual to be first transformed and then for the quantized transform coefficients to be entropy coded. By selecting transform and scan order based on intra prediction modes, higher coding efficiency is reporteedly achieved with small complexity increase. 2 transforms (pseudo DCT and DST) used, each for 4x4, 8x8, 16x16, implemented as matrix multiplications. HE: 3 scans (zig-zag, horizontal, vertical) – horizontal and vertical are transposed versions LC: adaptive scans (as from JCTVC-D304) 6 matrices (which are scalable versions of each other i.e. only one would need to be stored) per block size – different from HM quantization matrix. 2.1% BR reduction / 106% encoder time/ 105% decoder time for HE intra; 2.6% / 116% /114% for LC intra. Another report is given about a VLC modification which increases the gain to 3% for LC intra. 10.2.1.1.1.1.1.1.17JCTVC-D407 Verification of mode dependent intra residual coding [Seungwook Park, Jaehyun Lim, Byeongmoon Jeon] (missing prior, uploaded Thursday 20th, first day of meeting)
10.2.1.1.1.1.1.1.18JCTVC-D304 (m19070) CE7: Simplified MDDT scheme using symmetry-based scanning orders [H. Yang, J. Zhou, H. Yu] A SMDDT (Simplified MDDT) scheme was proposed in this document. It comprises two parts, ST (Simplified Transform) and SSO (Symmetry-based Scanning Order). In the SMDDT-ST, the combination of DCT and DST is used. In the SMDDT-SSO, 2 additional scanning orders as well as the existing zig-zag scanning order are used. The BD BR reduction of the proposed SMDDT scheme is 2.0% in the Intra HE case and 2.9% in the Intra LC case.
10.2.1.1.1.1.1.1.19JCTVC-D284 (m19049) CE7: Mode-dependent transform, residual reordering and coefficient scanning for intra prediction residue [Xin Zhao, Li Zhang, Siwei Ma, Wen Gao]
10.2.1.1.1.1.1.1.20JCTVC-D290 (m19056) CE7: Crosscheck of PKU’s proposal JCTVC-D284 by MediaTek [Xun Guo, Mei Guo] In this contribution, the results from Peking Univ. for Core Experiment 7 (CE7) on Alternative Transforms were provided, with both tool description and simulation results. Simulation results were obtained under the test conditions defined in CE7, and the detailed rate-distortion measurements using BD BR values were provided in accompanying spreadsheets.
Yüklə 7,54 Mb. Dostları ilə paylaş:
|