JVET-L0638 CE5-related: Retraining of context initialization values for CE5.1.4 [H. Kirchhoffer, C. Bartnik, P. Haase, S. Matlage, J. Stegemann, D. Marpe, H. Schwarz, T. Wiegand (HHI)] [late]
New context model initialization values are proposed for tests CE5.1.4.1 and CE5.1.4.2 (naming is according to the CE5 summary report JVET-L0025). Both tests use the state-based probability estimator of JVET-K0430 using an 8 and a 12 bit state variable per context model. CE5.1.4.1 doesn’t use custom window sizes, CE5.1.4.2 uses custom window sizes. Only the initialization values for the states are updated in this proposal (not the custom window size parameters). When comparing both CE tests (using new initialization values) to VTM-2.0.1 (using new initialization values), the advantage in Luma coding gain of CE5.1.4.2 (with custom window sizes) over CE5.1.4.1 (without custom window sizes) is only 0.06% for AI, 0.10% for RA, and 0.10% for LB.
Similar to results of JVET-L0552 in terms of VTM (though a different training procedure was used), and also unveils that the customized window sizes have less effect when better initialization is used. Conceptually, it is somehow obvious that one aspect of customized window sizes is that they only have effect if a correct customization is used.
CE6 related – Transforms and transform signalling (28)
Contributions in this category were discussed Monday 8 Oct 1720–2020 (chaired by JRO).