JVET-AL0179 Non-EE2: Intra Merge Mode Enhancements [M. Blestel, P. Andrivon (Ofinno), Y. Chang, V. Seregin, M. Karczewicz (Qualcomm)]
The method proposes to extend the Decoder-side Intra-Mode Derivation (DIMD) process by introducing the generation of an INTRA merge list. The INTRA merge list is built by reusing available information from spatial adjacent and non-adjacent neighbouring blocks with DIMD, TIMD, Intra Merge, and extended to Adaptive HoG for DIMD, OBIC, TIMD SAD, TIMD Merge, SGPM, TRML, and ITMP modes.
On top of ECM-16.1, the reported PSNR-Y, Cb, Cr, BD-rate and {EncT, DecT, EncVmPeak, DecVmPeak} results are as follows:
Information in the merge candidate list:
- Up to 8 intra-prediction modes, Intra-prediction weights, Intra-prediction location dependency (top, left or top-left), Block vectors, MTS transform types
The merge list size is 5 in this proposal. Template matching is performed to select the smallest cost to be further considered by the encoder, hence the encoder runtime increase is limited.
It was commented that the proposal blends up to 8 predictions, which seems to be excessive. It was commented by the proponent that this proposal uses the same process as used in DIMD, which also blends up to 8 predictions.
It was commented that complexity study should be performed. The adjacency map used in this proposal is from inter, and it should be studied what performance will be if adjacency map from intra (e.g. the one used by OBIC, or intraTMP, or IBC) is used.
It was commented that intra merge has been studied before. The difference between this proposal and the previous proposal is that this proposal uses available DIMD and TIMD information from more intra-predicted neighbouring blocks. Compared to previous intra merge proposal, this contribution has higher gain and less runtime increase.
It was agreed to study this in an EE. Performance should be tested using the current adjacency map as well as another adjacency map from an intra mode (e.g. OBIC, or intraTMP, or IBC).