JVET-W0110 AHG12: GPM with inter and intra prediction [Y. Kidani, H. Kato, K. Kawamura (KDDI)]
This contribution proposes geometric partitioning mode (GPM) with inter and intra prediction to enhance the coding performance beyond VVC. With the method, pre-defined intra prediction modes against geometric partitioning line can be selected in addition to motion vectors from a merge candidate list for each non-rectangular split region in the GPM-applied coding unit.
On top of the VTM-11.0+MCTF and ECM-1.0 which are defined as the anchor for tool-on and tool-off tests in EE2 of JVET-V2024, simulation results of the proposed method are reported as follows:
- Tool-on test (VTM-11.0+MCTF as the anchor):
- RA: BD-rate YUV: -0.18%, -0.51%, -0.36%; EncT: 102%; DecT: 101%
- LB: BD-rate YUV: -0.43%, -1.00%, -1.00%; EncT: 103%; DecT: 106%
- LP: BD-rate YUV: -1.01%, -1.84%, -1.81%; EncT: 111%; DecT: 101%
- Tool-off test (ECM-1.0 as the anchor):
- RA: BD-rate YUV: -0.04%, -0.29%, -0.18%; EncT: 101%; DecT: 100%
- LB: BD-rate YUV: -0.20%, -0.87%, -0.48%; EncT: 101%; DecT: 100%
- LP: BD-rate YUV: -0.38%, -0.93%, -0.92%; EncT: 107%; DecT: 101%
Intra prediction modes are selected such that their direction corresponds with the GEO separation line (parallel, perpendicular).
Question: Is the encoding time correct? The proponents assume it is due to the low number of pre-selected modes that are checked. For LP, the encoding time is more increased as in current CTC for LP GEO is not used at all (as it would require two MC ops).
Most likely, GEO itself contributes to the higher gain in LP.
Has it been tried with AI coding? No.
How is the MV handled in the partition that is intra coded? This needed to be checked.
It was agreed to investigate this in an EE.