JVET-AF0117 Non-EE2: extensions of IBC GPM [C. Ma, X. Xiu, W. Chen, H.-H.Jhu, C.-W. Kuo, N. Yan, X. Wang (Kwai)]

This contribution proposes to extend IBC GPM scheme in ECM10.0 from two aspects: First, IBC GPM with block vector difference (IBC GPM BVD) is introduced. Second, the split mode of IBC GPM is reordered with template matching methods. The simulation results on top of ECM-10.0 are summarized as follows:

1. IBC GPM BVD:

Screen Content:

AI:

Class F: {-0.20%, -0.12%, 0.01%, 103.4%, 99.1%, 100.0%}, Class TGM: {-0.26%, -0.24%, -0.23%, 103.3%, 99.8%, 100.0%};

RA:

Class F: {-0.10%, 0.04%, 0.02%, 102.3%, 99.8%, 99.9%}, Class TGM: {-0.21%, -0.14%, -0.17%, 101.4%, 99.4%, 100.0%};

Camera-view Content:

AI: {0.00%, 0.00%, 0.00%, 100.0%, 100.0%, 100.0%}, RA: {0.00%, 0.00%, 0.00%, 100.0%, 100.0%, 100.0%};

1. IBC GPM BVD + Split mode reordering:

Screen Content:

AI:

Class F: {-0.21%, -0.22%, -0.05%, 105.6%, 99.9%, 101.4%}, Class TGM: {-0.39%, -0.34%, -0.36%, 106.4%, 99.5%, 101.1%};

RA:

Class F: {-0.16%, 0.22%, -0.04%, 103.4%, 99.8%, 101.2%}, Class TGM: {-0.26%, -0.16%, -0.26%, 102.3%, 99.1%, 100.7%};

Camera-view Content:

AI: {0.00%, 0.00%, 0.00%, 100.0%, 100.0%, 100.0%}, RA: {0.00%, 0.00%, 0.00%, 100.0%, 100.0%, 100.0%};

It was commented that the split mode reordering seems to provide performance vs. runtime tradeoff that is less than IBC GPM BVD. The proponent said current encoder could be further optimized.

It was asked why there is no decoding time increase for split mode. This could be because only the first set is reordered, which is not always activated, therefore the impact on decoding time is minimal. However, encoder time increase due to split mode reordering is about 2%, which is likely due to additional RDO (currently a coarse RDO followed by a fine RDO) process in the current implementation being unoptimized.

Cross checker confirmed that results are matched, and suggests that the proposal is a straightforward extension. Upon examination of the code, the encoding time increase seems to be related to unoptimized code, and can be potentially reduced.

Several experts expressed interest in studying this proposal in an EE.

Currently in CTC for camera-view content, IBC GPM is not enabled, therefore this proposal has no impact on camera-view content performance. However, it might be interesting to test the proposal with IBC GPM turned on for camera-view content as the anchor, and see if there would be additional gain brought by this proposal. It could also be interesting to check the gain from IBC GPM on camera-view content in the latest ECM software. Proponents were requested to bring this as additional data to the next meeting.

It was agreed to investigate this in an EE, with separate EE tests on the two elements (IBC GPM BVD, and split mode reordering), and a reduction in encoder runtime was requested.