JVET-AL0150 Non-EE2: CABAC context switch of sig_coeff_flag and abs_level_gtx_flag for LFNST/NSPT [T. Kusakabe, K. Abe, T. Sugio, T. Nishi (Panasonic)]
In this contribution, it is proposed to introduce the process of switching the set of context models for CABAC coding of syntax elements sig_coeff_flag and abs_level_gtx_flag based on the prediction mode of the CU when LFNST/NSPT is used. In this proposal, a 2nd set of context models for LFNST/NSPT is added for CABAC coding of sig_coeff_flag and abs_level_gtx_flag, and selected based on the prediction mode of the CU. Two tests were conducted with different initial values for the context models for LFNST/NSPT. Test 1 used the same initial values for the new context models as the already existing context models for LFNST/NSPT in ECM-16.0, while Test 2 used initial values retrained with the CTC sequences for both the existing 1st set and the new additional 2nd set of context models for LFNST/NSPT. The following results were achieved compared to ECM-16.0:
Test 1: AI {Y 0.00%, U -0.02%, V 0.01%, EncT 100.2%, DecT 99.5%},
RA {Y -0.04%, U 0.04%, V 0.08%, EncT 99.7%, DecT 99.2%},
LDB {Y -0.01%, U -0.16%, U 0.16%, EncT 100.3%, DecT 99.4%}
Test 2: AI {Y -0.02%, U 0.01%, V 0.09%, EncT 100.7%, DecT 99.7%},
RA {Y -0.06%, U 0.01%, V 0.03%, EncT 100.0%, DecT 99.4%},
LDB {Y 0.01%, U 0.31% V 0.22%, EncT 99.9%, DecT 98.7%}
Basically, the second set has conceptually the same contexts as the first set, but they may adapt differently depending on the mode when they are used. Same initialization for both sets is used in test1 whereas the initialization of set 2 is retrained in test 2.
Number of additional context models would be 266.
There was some debate if the gain justifies the addition of new context models. It was commented that the gain by introducing mode-dependent switching for non-LFNST was higher.
Adding additional contexts is known to be a common way of getting somewhat more coding gain, but at this stage of exploration does not lead to substantial new insight.
No immediate aspect was identified to take action; this might be interesting for further study within a standards development.