JVET-AG0080 EE2-related: Extend block vector prediction for IntraTMP merge candidates [N. Qiu, J. Huo, Y. Ma, F. Yang (Xidian Univ.)]

EE2-1.7b, including EE2-1.2, EE2-1.3,EE2-1.4 and EE2-1.5.

Contribution JVET-AF0137(EE2-1.2) proposed extending the IntraTMP candidate list with all the merge candidates block vectors.

Contribution JVET-AF0226(EE2-1.3) proposed further including block-vector based prediction obtained from IntraTMP or IBC modes of the neighboring blocks in the candidate list of SGPM.

Contribution JVET-AF0275(EE2-1.4 and EE2-1.5) proposed IntraTMP extension to DIMD and LIC.

Contribution JVET-AF0129(EE2-1.8) proposed an auto relocated block vector prediction (AR-BVP) is introduced into IBC merge/AMVP candidate list construction.

This contribution merges both EE2-1.7b and EE2-1.8. In this contribution, an extension block vector prediction (EBVP) list is introduced into IntraTMP merge candidate list construction. EBVP can be derived as a guiding BV plus a BV of a reference block, located by the guiding BV.

On top of the ECM-11.0, simulation results of EE2-1.7b+EE2-1.8 are reported as below:

AI:

Class F: {-0.88%, -1.07%, -0.65% 105.4%, 106.6%, xx%};

Class TGM: {-1.63%, -1.51%, -1.54%, 103.2%, 106.7%, xx%};

On top of the ECM-11.0, simulation results of the proposed method are reported as below:

AI:

Class F: {-1.13%, -1.11%, -0.83% 106.7%, 108.3%, xx%};

Class TGM: {-2.22%, -2.08%, -2.08%, 104.4%, 107.9%, xx%};

The EBVP element of this proposal is similar to AR-BVP in EE2-1.8, but does not restrict the range of EBVP to be within the IBC search range (in comparison, the adopted AR-BVP has this restriction).

For the related contribution JVET-AG0063, the AR-BVP was also used without this restriction. The reason given by the proponents and confirmed by other experts was that, for IntraTMP merge candidates, the adjacent and non-adjacent candidates also do not have this restriction. It was suggested that restricting the BV range for all IntraTMP merge candidates (including adjacent, non-adjacent, and the proposed EBVP/AR-BVP) to be within IntraTMP search range should also be tested if we go for an EE test.

At the time of presentation, the proposal contains only results for screen content. It was reported by the proponents that although simulation for natural content is not complete, coding gain is expected.

The proponent of JVET-AG0063 confirmed that this proposal is indeed very similar, and coding gain for natural content is expected.

It was questioned why the encoder/decoder run time increased so much. The proponent said their run time was not accurate, the cross checker reported around 3% run time increase. This proposal also reported that EE2-1.7b+EE2-1.8 run time increase to be ~5%, which does not match the corresponding EE test results.

Presentation to be uploaded.

In the presentation of JVET-AG0080, the proponent showed the following luma gain on top of EE2-1.7b + EE2-1.8c in AI:

Class C: 0.07%, Class E: 0.04%, Class F: 0.25%, Class TGM: 0.61%

Other classes are not yet available at the time of presentation. During the discussion of JVET-AG0063, it was commented that the proposed method seemed to be a reasonable extension of the two EE tests that were adopted.

It was agreed to investigate JVET-AG0063 and JVET-AG0080 in the next round of EE. The following elements should be tested separately:

• The addition of AR-BVP/EBVP candidates
• Template matching based sorting of IntraTMP merge candidates, including performing template matching based sorting on a smaller number of maximum candidates (i.e. less than 50)
• Restricting the BV range for all IntraTMP merge candidates (including adjacent, non-adjacent, and the proposed AR-BVP/EBVP) to be within IBC search range