JVET-AM0159 Non-EE2: On ALF coefficient coding [I. Jumakulyyev, D. Bugdayci Sansli, J. Lainema (Nokia)]

In this contribution, two methods are proposed to improve ALF coefficient coding. Method 1 proposes to extend coding groups while method 2 proposes to re-assign the coding groups of coefficients.

The coding efficiency and runtimes of the proposed methods over ECM-17.0 (with Huffman frequency table re-trained) are reported as below {for Y, U, V, EncT, DecT}:

Method 1:

AI { -0.01%, 0.02%, -0.02%, 100.0%, 99.8%}

RA { -0.01%, -0.01%, -0.02%, 99.8%, 99.7%}

LDB { -0.02%, -0.15%, -0.17%, 99.8%, 99.8%}

Method 2:

AI { 0.00%, 0.01%, -0.01%, 99.8%, 99.3%}

RA { -0.01%, 0.03%, 0.02%, 99.7%, 100.2%}

LDB { -0.04%, -0.18%, -0.10%, 99.5%, 99.7%}

Discussion chaired by Y. Ye.

Method 1 extends the 2 groups of coefficients in the current ECM to 3 groups of coefficients.

Method 2 adjusts how the coefficients are assigned to each group, only 2 groups are used.

Huffman tables for each method are retrained.

It was commented that method 1 and method 2 cannot coexist, and one of them needs to be eventually selected, if we decide to change how coefficient grouping is done in ECM.

Proponent suggests to test both methods.

It was asked how much gain comes from Huffman table retraining. But method 1 requires new Huffman tables.

The proposal also reports that retraining of Huffman tables gives no gain on average (some classes have gain and others have loss).

Cross checker confirmed the results are matched.

It was commented that the reported results are not stable, with some sequences showing relatively large losses, esp. for chroma. It was commented that retraining the chroma ALF Huffman tables could alleviate the loss.

It was noted that the gains reported in this contribution are not under CTC, because the anchors have been changed to perform somewhat worse (with retrained Huffman tables).

Cross checker report contains comparison with ECM-17 without the retraining changes under CTC, and small gains were observed for RA and LDB for method 1, and LDB for method 2. Several experts expressed interest in further investigating the proposed methods, but some also expressed doubts about the benefits of this proposal.

Further study was recommended.