Joint Collaborative Team on Video Coding (jct-vc)
Block structures and partitioning
Yüklə 2,86 Mb.
|
5.9Block structures and partitioning5.9.1NSQT/AMPJCTVC-I0224 Results for Non-Square Intra Prediction with NSQT [X. Zhang, S. Liu, S. Lei (MediaTek)] This contribution reports methods and results for using non-square intra prediction with NSQT. The fundamentals are same as the methods proposed in AHG16 in the 8th JCT-VC meeting. The coding efficiency and run-time of the proposed methods built on top of HM6.0 software are reported. Experimental results report average 1.3% BD-BR reduction for all-Intra HE10 with encoding runtime increase of 31%. For random access HE10, experimental results reportedly show an average 0.6% BD-BR reduction with an encoding runtime increase of 4%. For low delay HE10, experimental results reportedly show an average 0.3% BD-BR reduction with an encoding runtime increase of 3%. The average decoding runtime variation is reportedly negligible. With further encoder complexity reduction, experimental results reportedly show an average 1.1% BD-BR reduction for all-Intra HE10 with an encoding runtime increase of 22% and negligible decoding runtime variation. It was asked what is the gain without non-square transforms, which would certainly be less, perhaps half. Concerns were raised that this is a non-trivial change, and it builds on NSQT which is still not fully stable yet. It should not be done at this late stage. It could potentially be interesting for future profiles. JCTVC-I0418 Cross-verification of Results for Non-Square Intra Prediction with NSQT (JCTVC-I0224) [R. Cohen (MERL)] JCTVC-I0323 Crosscheck of Non-Square Intra Prediction with NSQT (JCTVC-I0224) [Y. Chiu, W. Zhang, L. Xu, Y. Han (Intel)] [late] JCTVC-I0305 Test results on AMP and NSQT [X. Zheng (HiSilicon), L. Liu (HiSilicon), H. Yu (Huawei), I.-K. Kim (Samsung)] Nothing new technically – this document summarizes the performance for consideration in the profile/level discussion. JCTVC-I0517 Cross-check of JCTVC-I0305 [J. Xu (Microsoft)] [late] JCTVC-I0197 Cross verification of JCTVC-I0305 on NSQT from Huawei [E. Francois (Canon)] [late] JCTVC-I0306 Implementation of NSQT in HM [X. Zheng (HiSilicon)] [late] This contribution was for information, describing software, with no change relative to the status of the San Jose meeting. Note: The current text has bugs (tickets 505 and 506). It was suggested to provide a late input with proposed improved and bug free text on the basis of JCTVC-I0030.
The editor confirmed that the text is OK as it is suggested now; bugs have been fixed, and the text has become clearer. He also mentioned that the processes for luma and chroma RQT are becoming more divergent through NSQT. Decision: Adopt for draft 7.
This contribution provides an implicit TU split solution for asymmetric partitions when “QuadtreeTUMaxDepthInter” is set to 1 and NSQT is switched off. Experimental results reportedly show that the proposed solution contributes an average coding gain of 0.2% for RA-Main and RA-HE10, 0.4% for LDB-Main and LDB-HE10. Both encoder and decoder complexity are same as HM6.0. It was commented that this implicit splitting may be problematic with minimum transform size. It is also not clear whether all cases of AMP are covered. No support was expressed by other companies. JCTVC-I0452 Cross-check on implicit TU split prcoess for asymmetric partitions (JCTVC-I0308) [I.-K Kim (Samsung)] [late] JCTVC-I0150 Virtual large transform unit over maximum transform size with zero CBF [J. Kim, B. Jeon (LG)] This contribution proposes "virtual transform units" whose width and height are twice larger than the maximum transform size. When the virtual transform is applied, the luma CBF of the TU is derived as zero and the maximum size transform is applied to chroma components. It states 0.1%, −0.4%, −0.6% and 0.1%, −0.5%, −0.8% BD-BR, respectively, were measured in AI-Main and AI-HE10 conditions. It also states 0.1%, −0.6%, −0.6% and 0.1%, −0.5%, −0.5% BD-BR, respectively in RA-Main and RA-HE10 conditions. It states 0.1%, −0.7%, −0.9% and 0.1%, −0.9%, −1.0% BD-BR, respectively in the LB-Main and LB-HE10 conditions. One expert commented that the 32x32 was removed on purpose to reduce complexity (for the case of 4:2:0 and Main profile). The text does not have the restriction in general. No overall gain was obvious (loss in luma is observed). No action.
In this contribution, "virtual non-square transforms" whose size is larger than the maximum transform size are proposed. When the virtual non-square transform is applied, the luma CBF of the TU is derived as zero. It reportedly shows 0.0%, −0.2%, −0.1% gain for RA-HE10 and 0.1%, −0.4% −0.7% gain for LD-HE10 under the common conditions with maximum CU size of 64x64 and maximum TU size of 32x32. No action.
In this contribution, the transform split pattern of non-square transform TUs is restrained consistently regardless of the size of CU. Currently a TU in a CU larger than the maximum transform size is split into four square TUs until the TU size is equal to the maximum TU size. And then each square TU is split into four non-square TUs. In this contribution, a transform unit is always split into four non-square TUs regardless of the TU or CU sizes. It reports measured coding impacts of 0.0%, −0.1%, 0.0% for RA-HE10 and 0.1%, −0.1% −0.3% for LD-HE10 under the common conditions. No interest was expressed by other parties. There was claimed to be a simplification of syntax, but no draft text is provided. No action.
Yüklə 2,86 Mb. Dostları ilə paylaş:
|