JVET-AG0063 EE2-1.2 related: AR-BVP for IntraTMP merge candidates [L. Zhang, Y. Yu, F. Wang, H. Yu, D. Wang (OPPO)]

IntraTMP with merge candidates is investigated in EE2-1.2. This contribution proposes to add auto relocated block vector prediction (AR-BVP) to the IntraTMP merge candidate list to further improve the coding efficiency. Specifically, the AR-BVP candidates are inserted after the spatial candidates during the construction of IntraTMP merge candidate list.

On top of ECM11.0, simulation results of the proposed method + EE2-1.2 are reported as below:

The proposed method + EE2-1.2 over ECM-11.0 :

AI: { -0.13% Y, -0.10% U, -0.09% V, 101.0% EncT, 105.0% DecT } for Overall

{ -0.39% Y, -0.47% U, -0.10% V, 101.9% EncT, 105.2% DecT } for class F

{ -1.07% Y, -1.02% U, -0.99% V, 102.2% EncT, 109.4% DecT } for class TGM

The proposed method + EE2-1.2 over EE2-1.2 :

AI: { -0.07% Y, -0.06% U, -0.05% V, 99.7% EncT, 102.0% DecT } for Overall

{ -0.31% Y, -0.24% U, -0.24% V, 99.9% EncT, 101.5% DecT } for class F

{ -0.89% Y, -0.83% U, -0.80% V, 101.0% EncT, 106.5% DecT } for class TGM

RA results not complete when this contribution was presented.

This proposal adds template matching based sorting for the IntraTMP merge candidates, which is currently not part of ECM or EE2-1.2 (which was already adopted).

This proposal takes up to 50 candidates, and keeps 10 candidates after template matching-based sorting, the numbers of maximum candidates and kept candidates are the same as in EE2-1.2. It was commented that it would be desirable to test the performance with a smaller number of maximum candidates (i.e. less than 50) because this impacts the template matching-based sorting.

It was commented that the new aspects, i.e. template matching-based sorting and the addition of AR-BVP candidates, should be tested separately.

For AR-BVP that was tested in EE2-1.8, three variations were tested, i.e. trace path = 1, 2, and unlimited. The adopted variation was trace path = 1 given larger value of trace path did not provide much additional gain.

It was commented by one expert that given EE2-1.2 and EE2-1.8 were already adopted, this proposal seems like a natural extension of these two EE tests. There is some gain (0.07% in AI) for natural content, and more gain in screen content (0.31% for class F, 0.89% for class TGM, both in AI). The performance vs. complexity tradeoff seems to be reasonable.

JVET-AG0080 is a related contribution, with very similar method. See notes under JVET-AG0080.