JVET-D0064 Asymmetric Coding Units in QTBT [F. Le Léannec, T. Poirier, F. Urban (Technicolor)]
This contribution proposes a so-called Asymmetric CU coding tool on top of the Quad-Tree plus Binary Tree (QTBT) technology. The set of Binary Tree (BT) split modes of QTBT is extended, to support asymmetric binary splitting of a coding unit. As in QTBT, exactly one transform block and one prediction block is associated to each coding block.
Each coding stage (BT split signalling, prediction, transform, entropy coding, de-blocking filter) is adapted to new rectangular block sizes introduced in the JEM-3.0. These sizes correspond to a CU width or height multiple of 3, resulting from the division of a CU into 2 sub-CU with respective size ¼ and ¾ the size of the parent CU. In practice, this leads to a CU width or height potentially equal to 12 or 24.
To activate the Asymmetric CU tool in chroma components, the MaxBTSizeISliceC parameter is set to 64, instead of 16 in the test3 QTBT configuration chosen in 3rd JVET meeting. Around 6.5% average bitrate reduction in Chroma reportedly results from this modified MaxBTSizeISliceC, without any coding time increase.
For the asymmetric CU proposed tool, average coding gains of 1.54% in luma and around 8% in chroma are reported, compared to the unchanged JEM-3.0, with a 400% encoding runtime ratio. The decoding runtime impact is reported to be negligible. A faster version of the proposed tool is proposed with reported 1.16% gain in luma, around 7.5% in chroma, and 217% coding time ratio. Compared to the (JEM-3.0 + MaxBTSizeISliceC=64) configuration, 1.44% and 1.06% coding gains in luma are reported, respectively, with the slow and fast configurations of proposed tool.
Finally, it is suggested that the Asymmetric CU tool is integrated into a proposed dedicated EE, to be further investigated in JVET.
Increase of max BT size for chroma reportedly gives relevant gain (approx. 6% chroma rate reduction) and increases encoder run time only by 2% for AI configuration. For RA, the gain is probably lower but also no noticeable runtine increase would be expected (no results were available for that).
Decision (CTC): Adopt the change of QTBT MaxBTSizeISliceC to 64 (corresponding to 32 chroma samples).
The asymmetric partitioning provides interesting additional gain, but the encoder run time is increased in an unacceptable way in the current contribution. However, several experts expressed interest that this direction should further be investigated and there seems to be potential to further optimize encoder decisions.
Further study was recommended, in particular for the implementation for other configurations than AI and reduction of encoder runtime.
JVET-D0064 Asymmetric Coding Units in QTBT [F. Le Léannec, T. Poirier, F. Urban (Technicolor)]
Decision (CTC): Adopt the change of QTBT MaxBTSizeISliceC to 64 (corresponding to 32 chroma samples).