JVET-AF0084 Non-EE2: Update on IBC-LIC Model Merge mode [L. Zhang, Y. Yu, H. Yu, D. Wang (OPPO)]

This contribution proposes an IBC-LIC model merge mode that inherits IBC-LIC model parameters from spatial adjacent and non-adjacent neighbours, history candidates, or default models. The proposed method is implemented on top of ECM-10.0 software, and the performance is reported as follows:

AI: { -0.47% Y, -0.51% U, -0.36% V, 101.0% EncT, 100.6% DecT } for class F

{ -0.15% Y, -0.13% U, -0.11% V, 100.4% EncT, 102.6% DecT} for class TGM

RA: { -0.30% Y, -0.29% U, -0.23% V, 100.5% EncT, 100.0% DecT } for class F

{ -0.07% Y, -0.13% U, -0.10% V, 100.4% EncT, 101.8% DecT } for class TGM

It was asked if this method can be applied to IBC mode in camera-view content, the proponent commented that this had been tried on camera-view content and no gain was obtained. However, in the EE setting, results for camera-view content can be provided as well.

This proposal would require the model parameters from spatial adjacent and non-adjacent neighbours and history candidates to be additionally stored. Decoding time increase is likely due to more effort at the decoder to derive the IBC block vectors.

It was agreed to investigate this in an EE, also considering a potentially better tradeoff between performance and decoding time in terms of the number of inherited models.