JVET-O0425 Non-CE5: Supplementary results on alternative filter sets for Adaptive Loop Filter (JVET-O0090) [J. Taquet, P. Onno, C. Gisquet, G. Laroche (Canon)]

This contribution proposes to extend CE5-4 contribution by adding some limitations on the number of filters signalled by alternative luma filter sets, and by not using ALF Luma filters prediction from static filter sets. These limitations enables to reduce the amount of memory that a decoder should use to store the ALF filters coming from multiple APS.

In Test 1, the use of ALF Luma filters derivation from static filter sets is removed from VTM5.0.

In Test 2 to Test 5, the use of ALF filters derivation from static filter sets is removed from C5-4.3 test, and the maximum number of filters for each filter set is limited. In Test 2 and Test 4 the first filter set is limited to 7 filters, while in Test 3 and Test 5 it is limited to the maximum number of filters in VTM-5.0’s Luma filter set (i.e. 25 filters). The second filter set is limited to 3 filters in Test 2 and Test 3, and it is limited to 1 filter in Test 4 and Test 5.

The average Y/Cb/Cr BDR gains of Test 2 over VTM-5.0 are 0.10%/-0.91%/-1.27% for the AI configuration, 0.03%/-1.90%/-2.28% for the RA configuration, -0.11%/-2.38%/-2.66% for the Low Delay B configuration, and -0.22%/-2.53%/-2.67% for the Low Delay P configuration.

The average Y/Cb/Cr BDR gains of Test 4 over VTM-5.0 are 0.10%/-0.91%/-1.27% for the AI configuration, 0.05%/-1.92%/-2.28% for the RA configuration, -0.08%/-2.40%/-2.69% for the Low Delay B configuration, and -0.17%/-2.47%/-2.32% for the Low Delay P configuration.

The average Y/Cb/Cr BDR gains of Test 3 over VTM-5.0 are respectively 0.08%/-0.88%/-1.23% for the AI configuration, 0.03%/-1.90%/-2.29% for the RA configuration, -0.13%/-2.31%/-2.51% for the Low Delay B configuration, and -0.20%/-2.53%/-2.60% for the Low Delay P configuration.

The average Y/Cb/Cr BDR gains of Test 5 over VTM-5.0 are respectively 0.08%/-0.88%/-1.23% for the AI configuration, 0.05%/-1.92%/-2.29% for the RA configuration, -0.12%/-2.37%/-2.62% for the Low Delay B configuration, and -0.18%/-2.49%/-2.54% for the Low Delay P configuration.

In Test 2, Test 3, Test 4 and Test 5, compared to CE5-4.3 test, the memory for storing ALF Luma filters can be reduced by 80%, 44%, 84% and 48%, respectively. Compared to VTM5.0 considering all ALF filters, including Chroma, Test 2 and Test 4 enables to reduce the ALF filters’ memory by 45% and 53%, respectively, while Test 3 and Test 5 increases the ALF filters’ memory by 12% and 18% respectively.

One aspect (remove prediction from fixed filter set, Test 1) is subset of JVET-O0669

Other aspect was presented Sunday 1330

Tests 2-5 indicate that the results in CTC can be almost retained when reducing the number of filters. Different variants are tested.

Interesting information to assess the amount of storage necessary for storage of filter parameters.

Results also indicate that up to 25 filters may not always be needed.

Reasonable constraints for number per picture still need to be identified.

One possibility could be to constrain the amount of memory needed for decoded APS parameters per picture. See further discussion in plenary.