Back to Search Document details
17th Meeting: Brussels, January 2020 2019-12-31 20:10
Non-CE4: Triangular prediction mode with motion vector difference
Abstract
In this contribution, Triangular prediction mod with Motion Vector Difference (TMVD) is proposed. With TMVD, the MV of a triangular partition in TPM is refined as a sum of a MV derived from a merge candidate and a signalled MVD. The MVD is signalled in a similar way as MMVD, wherein indices of a direction and a distance are coded. Simulation results reported as below:
JVET-Q0315 Non-CE4: Triangular prediction mode with motion vector difference [K. Zhang, L. Zhang, H. Liu, Z. Deng, Y. Wang (Bytedance)]

In this contribution, Triangular prediction mod with Motion Vector Difference (TMVD) is proposed. With TMVD, the MV of a triangular partition in TPM is refined as a sum of a MV derived from a merge candidate and a signalled MVD. The MVD is signalled in a similar way as MMVD, wherein indices of a direction and a distance are coded. Simulation results reported as below:

RA: -0.18%, 103%, 98%

LB: -0.29%, 105%, 100%

Initially discussed in BoG JVET-Q0794.

Follow-up discussion in Track B Wed 15 Jan. (chaired by JRO):

The contribution reports interesting gain when combining MMVD with triangular mode. However, there are two issues:

- No results are available if it would still be effective with GEO

- It is more than just invoking the combination reusing existing logic. The proposal is also modifying the MMVD offset table.

No action is taken on this.

References:
JVET-N1010
Decisions
No action is taken on this.
Citation