International organisation for standardisation organisation internationale de normalisation



Yüklə 8,24 Mb.
səhifə163/203
tarix02.01.2022
ölçüsü8,24 Mb.
#15533
1   ...   159   160   161   162   163   164   165   166   ...   203

5.20.2Transform skipping


5.20.2.1.1.1.1.1.1JCTVC-H0141 Non-CE5: Improved transform skipping mode [J. Dong, Y. Ye (InterDigital)]

This contribution presented a modified design for the transform skipping mode (TSM) that was proposed in the 7th JCT-VC meeting (JCTVC-G575) and studied in core experiment (CE) 5 (JCTVC-G1205). First, the operation logic for TSM_1 and TSM_2 were modified with 16-bit implementation. Second, the number of arithmetic operations used in quantization and dequantization was reportedly reduced. Third, the look-up tables used for TSM to deal with a half-bit right shift were saved. This design was integrated into HM-5.0_tsm, which is the software base for CE5, and its R-D performance was tested under the common test conditions specified in JCTVC-G1200. The BD bit rate performance was reportedly a 0.0% to 0.1% increase on average for Classes A-E and 0.0% to 0.2% increase for Class F, compared with the baseline TSM. However, the proposal did not resolve the problems that were discussed in the context of CE5, and no action was taken on this.

5.20.2.1.1.1.1.1.2JCTVC-H0657 Non-CE5: A crosscheck report on complexity reduction for transform skipping mode (JCTVC-H0141) [R. Joshi (Qualcomm)] [late]
5.20.2.1.1.1.1.1.3JCTVC-H0061 Non Transform Mode for Inter Coding [Weijia Zhu, Wenpeng Ding, Yunhui Shi, Baocai Yin (BJUT)]

This contribution is related to the work of the AHG group on transform skipping, which addressed the transformation of residues. This contribution only focuses on the inter-residues situation. The inter- residues tend to be compactly represented in the DCT transform domain. But for some non-smooth residual blocks, transform skipping is a better choice. However, if all TUs adaptively select whether the 2D-DCT transform will be used according to RDO, the encoder complexity undoubtedly increases dramatically. It was reported that the major size of TUs using the proposed method is 4×4, so a scheme which only tests the proposed approach on the 4×4 TUs was proposed. This reportedly leads to a significant reduction of encoder complexity and a comparable coding performance. Simulation results reportedly show that the proposed scheme improves the coding performance of HEVC, especially for screen content sequences.

The proposal was to use only a full 2D transform or a full transform skip, for only 4x4 TUs in the inter coding case.

It was noted that a change of scaling in de-quantization can be implemented by shift.

The gain is 0.1-0.2% in common test conditions, and for Class F: 1.8% for RA, 3.7% for LD.

Encoding runtime increased 6% for RA.

No action was taken on this.

5.20.2.1.1.1.1.1.4JCTVC-H0325 Cross-check of BJUT proposal on transform skip mode (JCTVC-H0061) [H. Yang (Huawei)] [late]


5.20.2.1.1.1.1.1.5JCTVC-H0259 Efficient signalling of transform skipping mode [J. Kim, M. Kim (KAIST)] [late 02-04]

In this contribution, a signalling method of transform skipping modes for the non-square quadtree (NSQT) transform scheme was proposed. The proposed method for non-square transform skipping mode is to reduce the number of transform skipping modes by excluding the no-transform mode and to use a 1-bit flag to signal the transform skipping modes of 2D transform, 1D row transform and 1D column transform. The proposed method reduces the computational complexity by limiting the number of transform skip modes to be tested from 4 to 2 according to the non-square transform types. The encoding complexity of the proposed method is reportedly reduced to 2~3% over HM5.0_tsm in the common test conditions without degrading BD bit rate performance.

The proposal is to make signalling of skipping for the NSQT case dependent on the NSQT structure (e.g. skipping vertical is more likely when vertical size is low). It seemed questionable whether this is a complexity benefit (more likely an increase of dependencies). No compression gain was reported. No action was taken on this.


Yüklə 8,24 Mb.

Dostları ilə paylaş:
1   ...   159   160   161   162   163   164   165   166   ...   203




Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur ©muhaz.org 2024
rəhbərliyinə müraciət

gir | qeydiyyatdan keç
    Ana səhifə


yükləyin