JVET-AC0103 Non-EE2: AMVP mode with subblock-based temporal motion vector prediction [R.-L. Liao, J. Chen, Y. Ye, X. Li (Alibaba)]

This contribution proposes to apply SbTMVP to AMVP mode. Given a CU coded in AMVP with SbTMVP mode, the CU is predicted using a similar way to that of SbTMVP in merge mode except that the motion shift is signaled in the bitstream instead of deriving from left neighbouring block. The motion shift comprises a derived MVP and a signaled MVD, and the MVD only has 32 candidates which are composed of 8 motion directions and 4 motion magnitudes. It is reported that on top of ECM-7.0, the overall coding performance impact for {Y, U, V, EncT, DecT} is {-0.10%, -0.08%, -0.12%, 104%, 100%} for RA and {-0.28%, -0.49%, -0.53%, 104%, 101%} for LB.

Increase in encoding time due to additional motion search

Only one AMVP candidate is used.

It was agreed to investigate this in an EE. Reduction of encoding time should be targeted.

JVET-AC0213 is also targeting SbTMVP.