JVET-AG0126 AhG12: On ECM temporal partitioning prediction [G. Laroche, P. Onno (Canon)]

This contribution presents a temporal prediction method of the partitioning parameters and syntax elements. In this temporal prediction method, initially presented in the contributions JVET-AD0130, JVET-AE0132, the split modes allowances, the partitioning depths and the split syntax elements ordering are derived for each block according to the partitioning parameters of the current frame and parameters obtained from a temporal area. Compared to the previous contributions, the proposed method has been modified to increase the coding efficiency.

Compared to ECM-11.0, the average BDR gains and runtimes reported in this contribution are as follows:

-0.21% -0.14% -0.12% 98.3% 100.4% for the RA configuration,

-0.02% 0.04% -0.04% 96.0% 100.2% for the Low Delay B configuration.

It was asked if the reported runtime decrease was reliable. The proponent reported that the timing results were obtained from computing cloud with reliable timing information.

It was commented that the proposal requires some additional memory to store partition information of the collocated pictures, but the additional memory is relatively small compared to the motion information that already needs to be stored.

It was commented that the performance vs. complexity tradeoff is quite attractive, esp. given the encoding runtime reduction.

There are three main elements in the proposal: adaptation of maximum MTT depth, temporal prediction of split modes, and temporal prediction of split syntax. The gain of the maximum MTT depth element is estimated to be approximately 2/3 and complexity reduction, with the remaining gain coming from the other two elements.

It was commented that class F shows some loss in RA, and it was requested to provide results for screen content in the EE test.

It was asked how the QP of the collocated picture affects the temporal prediction. The proponent commented that prediction is more effective when the QP of the collocated picture is lower (i.e. better quality).

It was agreed to investigate this in the EE. Three elements of the proposal to be studied as follows:

• Temporal prediction of split modes
• Temporal prediction of split modes + Maximum MTT depth adaptation
• Temporal prediction of split modes + Maximum MTT depth adaptation + Prediction of split syntax element order

Class F and class TGM results are requested.