International organisation for standardisation organisation internationale de normalisation

5.18Transforms

5.18.1Core transform implementation

5.18.1.1.1.1.1.1.1JCTVC-H0305 The core spatial transform with fast algorithm [H. Zhu] [late 02-05]

Document was available 02-05, but no presenter was available.

5.18.2Alternative transforms

5.18.2.1.1.1.1.1.1JCTVC-H0456 Non CE 7: Experimental Results for the ROT [Z. Ma, F. Fernandes, E. Alshina, A. Alshin (Samsung)]

This contribution provided simulation results for the Rotational Transform (ROT) enabled in HM 5.0 against the HM 5.0 anchor results. The ROT was implemented using a lifting factorization with 5-bit multipliers and was applied to 8x8, 16x16 and 32x32 luma-component, intra residuals only. For the All-Intra (AI-HE) test case, the ROT provided a reported 0.6% performance gain with 142% encoder execution time relative to the anchor. Decoder execution times were reportedly identical to those of the anchor, within measurement-error margins, for all cases.

There was no need to present or discuss this contribution in detail; it had been submitted for information.

5.18.2.1.1.1.1.1.2JCTVC-H0559 Non-CE 7 : Recent results for secondary transforms for intra/inter prediction residual [A, Saxena, E. Alshina, F. Fernandes, T Nishi (Samsung), Y. Shibahara (Panasonic)]

Secondary transforms for intra and inter residue have been previously proposed in JCTVC-G108 and JCTVC-G632 respectively, and had been tested in Core Experiment 7. This contribution provided some results related to the proposals, JCTVC-H0125 and JCTVC-H0126 on secondary transforms. Specifically, it was reported that the transform schemes in JCTVC-H0125 and JCTVC-H0126 provide similar or additional gains when integrated with various other tools such as SDIP in AhG16, Transform Skip Mode in CE 5, and Data Hiding in CE 11. Also, it was reported that the complexity of the 4x4 secondary transform is negligible as compared to the 32x32 transform in HM, while it can provide more than half (roughly 60%-70%) of the gain for Intra TUs that a 32x32 transform is providing in HM.

It was reported that, when 32x32 transform is disabled and secondary transform in intra coding is enabled, coding losses are 0.4% and 0.2% in AI, for 8x8 and 4x4 secondary transform. Alternatively, gain could be obtained by keeping the 32x32 and adding the secondary transform.

It was also reported that the secondary transform retains the gain together with transform skipping.

This contribution was not a concrete proposal but rather various additional testing configurations using the same technology that was already reviewed in CE7.

5.18.2.1.1.1.1.1.3JCTVC-H0692 Non-CE7: Cross-check for Samsung Results for Secondary transforms in JCTVC-H0559 [A. Gabriellini, M. Mrak (BBC)] [late]
5.18.2.1.1.1.1.1.4JCTVC-H0689 Cross-check of JCTVC-H0559 on the combination of secondary transform and SDIP [H. Yang (Huawei)] [late]
5.18.2.1.1.1.1.1.5JCTVC-H0672 Non-CE7: Cross-check for Samsung Results for Secondary transforms in JCTVC-H0559 [J. An, X. Guo (MediaTek)] [late]
5.18.2.1.1.1.1.1.6JCTVC-H0603 Crosscheck of Samsung additional results JCTVC-H0559 for proposals JCTVC-H0125 and JCTVC-H0126 [F. Henry (Orange)] [late]
5.18.2.1.1.1.1.1.7JCTVC-H0685 Cross-verification of JCTVC-H0559 on boundary dependent transform [V. Seregin (Qualcomm)] [late]