Coding loss of LMCS with Kimono 420

[lienfei.chen]

LMCS has coding loss with Kimono420 test sequences on top of VTM-8.
The anchor is VTM-8 with LMCS on, and the Test is VTM-8 with LMCS off.
The coding performance for Kimono420 is shown as follows.
AI: 0.01% -0.27% -0.31%
RA: -2.53% 2.95% 6.61%
LB: -0.23% 1.34% 0.19%

When dual tree is DISABLE, the qualilty drop is also shown as follows
AI: -0.42% 1.70% 2.06%
RA: -3.03% 4.78% 9.71%
LB: -0.19% 1.83% 0.64%

[Michel0914]

Hi Lien-Fei,

Thank you for testing the Kimono sequence. We confirm your findings for the configuration settings you used. Note though that the VTM software allows LMCS parameters to be adjusted to improve LMCS performance on a sequence by sequence basis. We find that using the following parameter values produces gains (BDrate Y, U, V):
By setting LMCSOffset: 3, the LMCS gain is (-0.52%, 4.09%, 5.70%)
By setting LMCSOffset: 7, the LMCS gain is (-0.32%, 3.01%, 4.68%).

For both tests, we set:
LMCSUpdateCtrl: 0
LMCSAdpOption: 3
LMCSInitialCW: 1022

Note also that the current VTM software disables LMCS for sequences when a threshold number of pixels have luma values outside the standard range (16 <= luma <= 235, for 8-bit signals). For the top end of the standard range, the threshold for disabling LMCS is when the number of pixels having a value greater than 235 exceeds 0.03% of the total number of pixels in the picture. The Kimono sequence exceeds this threshold for a number of pictures in the second scene of the sequence. Thus, the Kimono sequence is on the edge of the LMCS enable/disable decision point as implemented in the current VTM encoder software.

Best Regards,
Fangjun