JVET-R0299 Additional fix for ALF virtual boundary processing [K. Andersson, J. Ström, Z. Zhang, J. Enhorn (Ericsson)]
At the last previous meeting, a low complexity fix for ALF virtual horizontal CTU boundary was adopted from JVET-Q0150. An alternative approach proposed in the same contribution was rejected since the increase of two luma line buffers and two chroma line buffers for each component was undesirable. This contribution proposes a combination of the two approaches in JVET-Q0150 as follows: Filtering of a row just above the virtual horizontal CTU boundary is performed as currently using the low complexity technique, i.e., not using samples below the virtual horizontal CTU boundary. When filtering a row just below the virtual horizontal CTU boundary on the other hand, this contribution proposes to change the filtering process so as to let it use also one row just above the virtual horizontal CTU boundary. This combined approach is asserted to further reduce visual artefacts from virtual horizontal CTU boundary processing. Proposal 1 of this contribution only changes the ALF filtering. It is claimed that the memory cost for proposal 1 is 60% of one 10-bit line buffers for luma samples and 60% of one 10-bit line buffer for each chroma channel for chroma samples. Proposal 2 of this contribution combines proposal 1 with an approach that avoids filtering across the virtual boundary also for SAO when filtering samples just below the virtual boundary, by employing padding. Samples just above the virtual boundary are SAO-filtered as currently. It is claimed that proposal 2 comes at no memory cost in terms of line buffers over the current draft of VVC.
The claimed benefit of the proposal is suppression of coding artefacts from virtual horizontal CTU boundary processing. The BD rate impact for luma for CTC is as follows:
Proposal 1: AI: -0.01%, RA: -0.03%, LDB: -0.09%
Proposal 2: AI: -0.01%, RA: -0.xx%. LDB: -0.xx%
Similar encoding and decoding times as the anchor are reported.
It is claimed that the method improves over the method from JVET-Q0150 adopted by last meeting, but requires approx. 0.6 additional line buffer (by using buffer jointly with SAO).
It is however pointed out that SAO does not need to store sample values, so it would be more like 1 line buffer.
The proposal would require a substantial amount of changes, and the additional subjective benefit over the JVET-Q0150 method may not be too large.
No action was taken on this.