JVET-AE0105 Non-EE2: Local illumination compensation with multiple templates [Y. Wang, K. Zhang, Y. He, H. Liu, L. Zhang (Bytedance)]
This contribution presents a method of local illumination compensation with multiple templates. Apart from using a template comprising neighbouring samples both top to and left to the current block, two new LIC modes using top-only or left-only template are proposed to derive the parameters of LIC. On top of ECM-9.0, simulation results of the proposed method are reported (estimated) as below:
- RA: {-0.09%%, -0.20%, -0.16%; 109.4%, 100.1%};
- LDB*: {-0.06%, -0.15%, -0.32%; 117.6%, 99.6%}.
Different shapes could be used for different motion vectors in same location.
Is there dependency of templates between adjacent blocks? Not currently exploited.
Possibility to reduce encoding time? No concrete ideas or preliminary experimentation were available for this yet.
This is also applied to chroma.
Several experts expressed this is an interesting approach. Further study was recommended on largely reducing encoding time while keeping the benefit in compression.