JVET-AC0146 AHG12: Filtered Template Matching based Intra Prediction (FTMP) [R. G. Youvalari, D. Bugdayci Sansli, P. Astola, J. Lainema (Nokia)]

This contribution presents a template-matching based intra prediction mode for luma blocks, which is called Filtered Template Matching based intra Prediction (FTMP). Prediction is generated by applying a 6-tap filter on the reference block that is obtained from template-matching based search. The filter input consists of 5 spatial luma samples and a bias term. Filter coefficients are derived via regression based MSE minimization on reconstructed samples over the template areas of the current and reference blocks. The mode is signaled with a flag conditioned on intra TMP flag. The impact on coding efficiency and runtimes over ECM-7.0 is reportedly {for Y, U, V, EncT, DecT}:

AI { -0.14%, -0.15%, -0.19%, 105%, 101%}, RA { -0.04%, -0.04%, -0.11%, 101%, 100%}

It was commented that the results in the abstract refer to using 8 candidates; in the presentation, also results with less candidates are shown, but no encoder run times for those – it is likely that the run times are lower. Higher number of candidates also increases decoder complexity.

It was further commented that using more candidates is having a similar effect as fusion methods.

Is the method always applied? Signalled by flag, RDO check at encoder.

A similar approach is in the proposal JVET-AC0109, but different method of deriving the filtering, and different encoder decision. Main difference seems to be number of candidates. The proposal JVET-AC0109 seems to have a better tradeoff.

It was requested to communicate offline with the proponents of JVET-AC0109 to identify what the new aspects of this proposal are, and if that would be relevant to investigate in the EE. If possible, that should use the same codebase.