JVET-R0322 CCALF virtual boundary issue for 4:4:4 and 4:2:2 format [X.W. Meng (PKU), X. Zheng (DJI), S.S. Wang, S.W. Ma (PKU)]
In this contribution, the CCALF virtual boundary issue for 4:4:4 and 4:2:2 format is addressed. Specifically, the current processing chroma sample and its collocated luma sample may belong to different Coding Virtual Blocks (CVB, the block between two neighbouring horizontal virtual boundaries) for 4:2:2 and 4:4:4 chroma format, which means that these chroma samples have to be stored until deblocking and SAO of collocated luma samples are done. This will cause extra hardware overhead.
In this proposal, a bug in CCALF covariance calculation process in VTM-8.0 is reported and fixed, firstly. Then, three methods are proposed to solve the virtual boundary issue.
Aspect 1: Bug fix of VTM-8.0 (YUV422 natural sequances)
- AI: 0.00% 0.00% 0.00%;
- RA: 0.00% -0.09% -0.01%;
- LDB: 0.02% -0.01% 0.07%
Aspect 2: Solution1 (Disable CCALF when a chroma sample and its collocated luma sample belong to different CVBs.)
- YUV444:
AI: 0.00% 0.01% 0.07%; RA: 0.04% -0.04% 0.18%; LDB 0.03% -0.09% -0.01% - YUV422:
AI: 0.00% 0.04% 0.09%; RA: 0.00% 0.06% 0.07%; LDB: 0.01% 0.07% 0.12%
Aspect 2: Solution 2 (For 4:2:2 and 4:4:4 chroma format, a chroma sample and its reference luma sample are forced to belong to the same CVB)
- YUV444:
AI: 0.00% 0.01% 0.05%; RA: 0.05% -0.01% 0.02%; LDB 0.02% -0.12% -0.06% - YUV422: AI: 0.00% 0.05% 0.08%; RA: 0.00% 0.04% 0.13%;
Aspect 2: Solution3 (For 4:2:2 and 4:4:4 chroma format, the location of chroma VB is changed from 2 lines to 4 lines above a CTU boundary)
- YUV444:
AI: 0.00% -0.02% -0.04%; RA: 0.01% -0.07% 0.03%; LDB 0.04% -0.04% -0.16% - YUV422:
- AI: 0.00% 0.00% -0.02%; RA: 0.01% -0.11% -0.08%; LDB 0.00% -0.15% 0.01%
Was presented in track B Sun 19 Apr. 0810 (chaired by JRO)
Aspect 1 is an obvious bug in VTM8 encoder (not filed as ticket so far)
Aspect 2: Same solutions as in JVET-R0233 – see notes there.
Decision (BF/SW): Adopt JVET-R0322 aspect 1