International organisation for standardisation

5.8.5Subtest 3: Memory reduction in ALF decoding

5.8.5.1.1.1.1.1.1JCTVC-F272 CE8 subtest 3: Line memory reduction for in-loop filtering [S. Esenlik, M. Narroschke, T. Wedi (Panasonic)]

The proposal reduces line memory requirements by using partially deblocked (only horizontally deblocked) samples at the LCU boundaries. In other words, at the bottom LCU boundary, the available partially deblocked samples are used as the input to SAO and ALF, instead of the fully-deblocked samples (horizontally and vertically deblocked), for which waiting and thus line memory would be necessary.

It was noted that there is a non-CE proposal that is related (JCTVC-F054).

This somewhat changes the prior intent of "parallel deblocking filtering" – changing the data flow between the deblocking filter and ALF processes.

The coding efficiency impact is basically negligible in PSNR terms.

It was asserted that this proposal improves subjective quality related to SAO. (There are also other proposals for fixing that issue.)

Regarding the proposed draft text: At first glance, the text appears to resemble an intended method of implementation rather than simply a specification of mathematical relationships.

Regarding the software: This introduce substantial changes in the software, introducing a line buffering scheme. It sounds as if the offered software differs from what we might want for reference software – i.e., readability, flexibility, and the ability to easily test alternative approaches.

The scheme seems to have technical merit, but the text and software do not seem mature.

No action, pending improved text and software (with review by software coordinators and text editors).

JCTVC-F054 is a somewhat competing proposal.

The group decided to set up a CE to compare JCTVC-F272 with JCTVC-F054 (which should include subjective study).

5.8.5.1.1.1.1.1.2JCTVC-F521 CE8 Subset 3: Cross-Verification of Panasonic’s adaptive loop filter (JCTVC-F272) by Qualcomm [I. S. Chong, M. Karczewicz (Qualcomm)]

Software was studied as well as executed.