6.1.4Transforms and transform coefficient coding
JCTVC-M0098 AHG5: Independent chroma transform depth from luma transform depth for non 4:2:0 format [J. Kim (LG??)]
(Discussion chaired by D. Flynn.)
In this contribution, the chroma residual quadtree could be decided independently from the luma tree for a given CU by signalling transform split flags for the chroma component. To reduce the increase in encoding time to a mere 10% and reduce the bits to encode chroma transform split flag, this contribution also proposes to limit chroma RQT depth to that of luma. It reports gains of 0.1%, 0.4%, 0.6% for AIMT, 0.1%, 0.4%, 0.4% for AIHT and 0.1%, 0.3%, 0.3% for AIST. Also it reports 0.9%, −0.6%, −0.3% and 0.9%, −0.7%, −0.5% for RAMT and RAHT respectively and 1.2%, −1.3%, −1.0% and 1.2%, −1.1%, −1.3% for LBMT and LBHT respectively.
For reference, results are also presented for the case when the chroma RQT is fully independent from luma, where, for a 30% increase in encoder run time, it attains average gains of 0.0%, 0.6%, 0.7% in the case of AIMT, 0.0%, 0.6%, 0.5% in the case of AIHT and 0.1%, 0.5%, 0.4% in the case of AISHT. It also reports 0.4%, 0.1%, 0.2% and 0.5%, 0.0%, 0.1% for RAMT and RAHT respectively and 0.5%, −0.5%, −0.3% and 0.6%, −0.3%, −0.4% for LBMT and LBHT respectively.
The mode decision is implemented by performing an RDO search, where for each CU the luma RQT tree is decided first followed by chroma. It was pointed out that the HM currently includes other features that trade encoder runtime for performance increases, such as the joint luma-chroma search where the chroma modes are pre-estimated and used to influence the luma decision. No comparison to such techniques was provided.
No action.
Dostları ilə paylaş: |
