JVET-P0551 AHG16/Non-CE5: On ALF boundary padding [N. Hu, V. Seregin, M. Karczewicz (Qualcomm)]
Similar topic as to JVET-P0053 and JVET-P0157.
In VVC draft 6, there are two ways to handle unavailable pixels beyond boundaries for adaptive loop filter (ALF).
- Repetitive padding: for picture boundaries and 360° virtual boundaries that are not on CTU boundaries, pixels are padded towards outside of the boundaries. Then after padding, Laplacian calculation and filtering process are performed without changes.
- Mirrored padding: for the boundaries (virtual boundary for line buffer reduction, sub-picture boundaries and 360° virtual boundaries on CTU boundaries), the samples located beyond the line buffer boundary are excluded in ALF process, where Laplacian is re-scaled and truncated symmetrical filtering is applied.
In JVET-N0654, it was proposed to unify the padding methods, such that only one padding method is applied to all type of boundaries. In version 2, results are updated.
Three possible unification and simplification by reusing existed methods are applied:
- apply mirrored padding to 360 virtual boundaries and picture boundaries
- apply repetitive padding for all boundaries excluding virtual boundaries for ALF line buffer reduction
- apply repetitive padding for all boundaries
Method 2 is the same as method 4 in JVET-P0157 and first proposal in JVET-P0053.
BoG recommendation: Adopt method 2 in JVET-P0551, which is also method 4 in JVET-P0157 and the first aspect of JVET-P0053. The software is from JVET-P0551. (Clarification discussed on Thursday, Oct. 10, at 18:20.)
For cases of filtering with repetitive padding, the classification padding should be the same as currently in the cases of repetitive padding (e.g. as at the picture boundary). ALF virtual boundary classification should not be changed.