JVET-R0164 AHG10: Mean-scaled SATD for VTM encoder [J. Lainema, A. Hallapuro (Nokia)]
This contribution reports that a coding efficiency improvement is achieved by scaling down the DC coefficients in the SATD (Sum of Absolute Transformed Differences) calculations performed by the VTM encoder. It is proposed to apply a weight of one quarter for the absolute value of the DC coefficient resulting from the Hadamard transform of the SATD process, while keep applying a weight of one for the rest of the coefficients.
Coding efficiency impact is reportedly on average -0.12 %, -0.29 % and -0.37 % for AI, RA and LD-B configurations, respectively. It is further reported that the gains in coding efficiency tend to be larger for the higher resolution categories. For example, in the case of Class A1, the impact in RA configuration is reported as -0.57 %.
[v2 added LD-P results and a new section on output image analysis]
Reasonble gain which comes practically for free by a simple encoder change.
Decision (SW/CTC): Adopt JVET-R0164, modify SATD cost function both fo ME and initial intra mode preselection by giving less weight to DC coeff. This should be made configurable by macro.
It would be desirable to have the same change implemented in HM. It was reported that the proponents already checked that and got almost the same gain for intra, but lower gain for RA (might however not yet be fully correct implementation in HM)