JVET-N0489 CE1: Bilateral filter tests [J. Ström, P. Wennersten, J. Enhorn (Ericsson), K. Reuze, D. Rusanovskyy, M. Karczewicz (Qualcomm)]
This document describes the core experiment tests CE1-1.1 through CE1-1.7 related to bilateral filtering. Test tests span different placements of the filter; as a post-reconstruction filter (tests CE1-1.1 through CE1-1.4, and CE1-1.7) as well as a loop filter (CE1-1.5). It also spans different amounts of precision in the calculations (CE1-1.1 versus CE1-1.2) and parametrization aspects (CE1-1.7).
CE1-1 tests use a bilateral filter as a post-reconstruction filter. Hence after the reconstruction is formed by adding the prediction to the residual and clipping, the filter is applied. To filter the intensity of the current sample, the surrounding four neighbours (above) (below) (left) (right) are used.
Samples used in the weighted sum.
The filtered value is then calculated as a weighted sum of the surrounding delta-intensity values as
where the deltas are the difference against the center sample, etc. The weights are calculated as
Here, and depend on the qp and block size, and is a non-local average of absolute differences in a 3×3 neighbourhood as described in JVET-J0021 section 4.
By varying the amount of shift, i.e., the variable in Equation 5.1.1, it is possible to choose small values for . This means that the multiplication in Equation 5.1.2 can be implemented as a two-bit multiplier plus an addition. Hence, compared to the method presented in Macau, the two multipliers have changed the following way:
- 9 bit x 8 bit multiplier is now 2 bit x 8 bit
- 12 bit x 11 bit multiplier is now 10 bit x 11 bit
For more information, see JVET-N0489 Section 2.