JVET-N0144 Non-CE1: Simplification of division calculation in local illumination compensation [Y. Yasugi, T. Ikai (Sharp)]
This contribution proposes a method to reduce the table size in scale value derivation in local illumination compensation (LIC). In CE4, the LIC scale value derivation is division free with the 16-bit inverse value multiplication. However the multiplication requires an operation with up to 28 bits and a table of 1024 bits. The proposed method reduces the bit width to 10 bits and the table size to 64 bits. The experimental result reportedly show the luma BD rate performance on average are -0.42% and -0.39% for RA and LB configuration, respectively.
The algorithm of the proposed method is basically as same as the scaling method in CCLM of VTM-4.0, which was originally proposed in JVET-M0064.
The method corresponds to the “1-5.1a” results of CE1 which is not the desirable approach. Relative to that, the loss is 0.04%.
Table has 4 bit instead of three in CCLM.
The unification with CCLM is probably going into the right direction, but here only one part of it is used. Furthermore, currently LIC does not seem to have enough benefit for making it a candidate for VVC (as per results of CE1).