JVET-AG0078 AHG12: Intra-prediction using Merged Histogram of Gradients [S. Blasi, I. Zupancic, P. Astola, J. Lainema (Nokia)]

This contribution proposes a new intra-prediction mode based on the computation of a Merged Histogram of Gradients using information from neighbouring blocks. The Merged Histogram of Gradients is then used to derive a set of intra modes and weights that are blended together to compute the intra prediction for a block. The impact on coding efficiency and runtimes of the proposed method over ECM-11.0 are reportedly {for Y, U, V, EncT, DecT, EncVmPeak, DecVmPeak}:

AI {-0.10%, 0.00%, 0.00%, 100.6%, 99.6%, 100.4%, 100.0%}

RA {-0.04%, 0.02%, -0.03%, 99.6%, 100.7%, 99.8%, 100.4%}proposal shows interesting gain especially for 4K sequences, with limited impact on run time. The proponent mentioned that aside from normative restriction on the use of the new mode (block size, availability of MIMD and DIMD neighbor), there is also some fast encoding algorithm in the proposal that achieves better performance vs. complexity tradeoff than bruteforce RDO.

Compared to DIMD merge, this proposal includes a new algorithm that composes a synthetic histogram of gradients based on neighboring blocks (which do not have to be DIMD blocks). Both adjacent and non-adjacent neighboring blocks are used.

It was agreed to investigate this in the next round of the EE. Results are requested for using the normative restriction as proposed, and for removing the normative restriction.