JVET-L0552 Training initial CABAC states [F. Bossen (Sharp)] [late]

Initial CABAC states are retrained on the data set defined by common SDR test conditions. A BD rate improvement in the order of 0.2% on average for RA, LB and LP is reported using such retrained initial states. For class D, average gains in excess of 0.5% are reported. Additional experiments include retraining only subsets of the initial states, and recursively retraining a second time. It is suggested to further study the impact of such training on the development of the VVC standard.

The overall gain (luma) was reported as 0.04%, 0.18%, 0.23%, 0.20% for AI, RA, LDB, and LDP, respectively.

When training only contexts related to coefficients, almost no improvements. This probably explains that the gain for intra is lower.

The effect of retraining is larger when applied to CE5 (fixed window size), but similar to VTM for CE5 (with adaptive window size).

A weighted method was used, such that the results are independent from image size.

It is discussed if it may be beneficial to use training of CABAC initialization regularly with each new VTM to prohibit contributions from using trained elements.

The proponents of new syntax elements likely use some way of training anyway.

It was commented that if initialization is trained on the current test set, that test set should not be used for verification testing of the final standard.

To prevent overtraining, it could also be investigated if the trained initialization is still beneficial for an independent set of sequences (and there are enough of them available). This could for example be done by comparing the previous set of initializations against the new one with an independent set (sanity check).

The following suggestion was made: to retrain the initializations with the CTC set for each new major version of the VTM; verify with an independent set of sequences (tbd), which should be large enough that the training is not overtrained (taking the initializations before and after, and see that the deviation of results is not severe on that other set).

This was agreed in the plenary Wed. 10 Oct. 1400.