JVET-AC0212 Non-EE2: Template matching for IBC BVD suffix derivation [P. Nikitin, Z. Zhang, H. Huang, V. Seregin, M. Karczewicz (Qualcomm)] [late]

This contribution proposes BVD suffix derived using template matching search, and a flag is signalled to indicate whether the suffix is transmitted or equal to zero.

AI: -0.29% (Y), -0.10% (U), -0.37% (V), 99% (EncT), 102% (DecT) for class F

-0.68% (Y), -0.63% (U), -0.63% (V), 103% (EncT), 102% (DecT) for class TGM

RA: -0.27% (Y), -0.17% (U), -0.25% (V), 94% (EncT), 96% (DecT) for class F

-0.39% (Y), -0.45% (U), -0.28% (V), 102% (EncT), 102% (DecT) for class TGM

RDO search at encoder? No.

How many template matching operations are performed? Depends on size of prefix, could become large when prefix is large (due to the case that BVD is typically rather small, at least for class F). This might be the reason why for some sequences runtime is significantly larger. Worst case complexity might be rather large.

How would that combine with new adoptions, in particular sign prediction? Might not be straightforward. One expert comments that according to his opinion gain might no longer be achievable on top of new ECM adoptions. This is however a case that always may happen, and in cases when the gain disappears ontop of new ECM, technology investigated in an EE would not be adopted, anyway.

It was agreed to investigate this in an EE. Beyond the best way of combining with new adoped proposals, also methods should be investigated to avoid excessively large search ranges.