Back to Search Document details
16th Meeting: Geneva, October 2019 2019-09-24 13:53
Non-CE6: Context modeling for LFNST index coding
Authors: T. Tsukuba Sony M. Ikeda Sony Y. Yagasaki Sony Teruhiko Suzuki Sony Takeshi Tsukuba
Abstract
This contribution proposes a simplified context modeling for LFNST index coding by removing treeType -dependent context assignments to reduce the number of context variables; The first bin of LFNST index is context-coded with single context variable and the second bin is bypassed.
Meeting Notes
JVET-P0271 Non-CE6: Context modelling for LFNST index coding [T. Tsukuba, M. Ikeda, Y. Yagasaki, T. Suzuki (Sony)]

In this document a bugfix for the configuration parameter MTSIntraMaxCand is proposed, which by itself is claimed to have no impact on the CTC. Furthermore, it is proposed to set the CTC value of the parameter for AI and RA to from currently 3 to 1, avoiding testing transform combinations which are reportedly rarely used. A minor BD-rate impact on the CTC results is reported, namely 0.00% (Y), 0.00% (U), 0.01% (V) for AI and 0.01% (Y), 0.01% (U), -0.01% (V) for RA.

The mismatch between SW and text is already resolved (ticket issued). Proposal to reduce the number of context variables of bin 0 from 2 to 1, removes dependency on tree type.

Very small loss (0.01% in RA) was observed unless retraining the matrices, and the complexity advantage appears negligible. No action was taken on this.

References:
JVET-N1010 JVET-N1021 JVET-O0368 JVET-O2001
Cited:
JVET-P0026 JVET-P0771 X
ARCHIVE v1 :
JVET-P0271.docx JVET-P0271_Test2_CTC.xlsm JVET_P0271_Test1_CTC.xlsm
Decisions
Very small loss (0.01% in RA) was observed unless retraining the matrices, and the complexity advantage appears negligible. No action was taken on this.
Citation