Back to Search Document details
20th Meeting: by teleconference, October 2020 2020-10-01 09:38
AHG12: Rice parameter derivation for high bit-depth coding
Abstract
This contribution proposes to improve the derivation of rice parameter for regular residual coding (RRC) for high bit-depth coding. VVC derives rice parameter in RRC with look-up table and its range is restricted from 0 to 3. It is asserted that this restriction causes worse performance of VTM in high bit-depth coding compared to HEVC Format Range extension profile where rice parameter derivation can be extended when persistent_rice_adaptation_enabled_flag is equal to 1. To resolve this problem, this contribution introduces a rice parameter derivation method using a predicted remainder value.
JVET-T0085 AHG12: Rice parameter derivation for high bit-depth coding [T. Hashimoto, T. Ikai (Sharp)]

This contribution was discussed in JVET session 11 at 2240 on Monday 12 October (chaired by GJS & JRO).

This contribution proposes to improve the derivation of rice parameter for regular residual coding (RRC) for high bit-depth coding. VVC derives Rice parameter in RRC with look-up table and its range is restricted from 0 to 3. It is asserted that this restriction causes worse performance of VTM in high bit-depth coding compared to HEVC Format Range extension profile where Rice parameter derivation can be extended when persistent_rice_adaptation_enabled_flag is equal to 1. To resolve this problem, this contribution introduces a Rice parameter derivation method using a predicted remainder value.

It is reported 1.52% and 25.02% in bdrateY in AI-12 qp= −13, −8, −3, 2, 7, and AI-16 −33, −28, −23, 18, 13, and 0.24% and 16.95% in bdrateY in LDB-12 qp= −13, −8, −3, 2, 7, and LDB-16 −33, −28, −23, 18, 13 respectively.

This is conceptually similar with JVET-T0072, but uses a modified version of the current lookup table for Rice parameter derivation. This avoids the latency problem of JVET-T0072 but requires a second table to be used depending on bit depth.

The contributor said this scheme is consistent with the current design in VVC v1.

A difference is that this scheme does not use the current LUT approach and instead uses a formula.

(A previous editorial proposal for VVC v1 had a formula approach that produces the same results as the current LUT approach, so using a formula is possible with VVC v1.)

It was asked whether this has a backward compatibility issue. The proponent suggested having a new flag to select between the two methods.

JVET-T0072 has a history tracking, basically as in HEVC range extensions. This one does not have that.

It was commented that it would be good to study more test cases before selecting one.

Decisions
It was commented that it would be good to study more test cases before selecting one.
Citation