JVET-V0104 EE2-related: TU-level adaptive self-guided filter [W. Yin, K. Zhang, L. Zhang, Y. Wang, H. Liu (Bytedance)]
This contribution presents a TU-level adaptive self-guided filter (TU-ASGF) with five taps and a cross-shape. TU-ASGF is applied on the reconstruction samples of a TU. The filtering strength is based on the TU size, QP, and statistical information adaptively.
On top of the EE code of JVET-U0100, simulation results of the proposed method are reported as below:
Tool-on test (VTM-11 as the anchor):
AI: -0.42%, 122%, 123%; RA: -0.32%, 112%, 111%; LB: -0.41%, 108%, 113%.
Tool-off test (EE code as the anchor):
AI: -0.43%, 115%, 110%; RA: -0.24%, 104%, 101%; LB: -0.28%, 104%, 102%.
On top of the EE code of JVET-V0094 (bilateral filter), simulation results of the proposed method are reported as below:
Tool-on test (VTM-11 as the anchor):
AI: -0.71%, 126%, 122%; RA: -0.55%, 110%, 108%; LB: -0.64%, 107%, 110%.
Tool-off test (EE-bilateral filter code as the anchor):
AI: -0.22%, 115%, 114%; RA: -0.09%, 106%, 104%; LB: -0.11%, 103%, 103%.
The results indicate that there is an overlap in gain with the bilateral filter (concepts are similar, anyway)
The filter is however in the intra prediction loop whereas the bilateral filter is (currently) a loop filter.
Which TU sizes are filtered? Minimum 64 samples.
Are intra blocks filtered? Yes.
In case of neighbouring blocks being intra blocks, samples from these are filtered.
When the bilateral filter was originally introduced, it was also operating in the prediction loop and had higher than currently as a post filter. Then, restrictions were imposed, such as not using it for intra coded blocks, not using samples from neighbouring blocks, and the gain went down.
A question would be if the self guided filter would still give additional gain when similar restrictions are imposed as for the bilateral filter. The most consequential way would be operating it as a loop filter
Proponents are asked to perform further study if the self-guided filter could be beneficial in addition to the bilateral filter. Proponents were initially asked to investigate if some action regarding EE could be taken from this meeting.
In version 3, results with a variant which implements the filter as loop filter are reported (only AI and for some classes). The gain is substantially decreased.
Further study (not CE) is recommended, if there is potential for more gain.