JVET-AO0175 Non-EE2: Scharr Filter for Rectangular blocks [A. C. Sidiya, S. Deshpande (Sharp)]
It is proposed to use the Scharr filter instead of the current Sobel filter for HOG-based gradient computation in ECM. It is asserted that Scharr filter reduces directional bias in discrete gradient estimation, offering improved rotational symmetry over Sobel filter.. A normalization factor is applied with the proposed Scharr filter to maintain a gradient dynamic range comparable to Sobel filter, ensuring compatibility with existing thresholds and histogram binning.
On top of ECM19.0, simulation results of the proposed method are reported as follows:
For AI configuration: -0.01%, 0.01%, -0.01%, EncT and DecT not reliable.
The (very small) gains are consistent across all classes, as shown below.
All Intra Main 10 | |||||||
Over ECM-190 | |||||||
Y | U | V | EncT | DecT | EncVmPeak | DecVmPeak | |
Class A1 | -0.01% | 0.02% | -0.01% | 97.8% | 99.0% | 100.1% | 100.0% |
Class A2 | -0.02% | 0.10% | 0.01% | 97.7% | 98.5% | 100.1% | 99.9% |
Class B | 0.00% | -0.01% | -0.06% | 97.5% | 99.1% | 100.0% | 100.0% |
Class C | -0.02% | -0.06% | -0.03% | 97.5% | 99.0% | 100.4% | 100.0% |
Class E | -0.01% | 0.02% | 0.11% | 97.6% | 99.8% | 99.0% | 99.9% |
Overall | -0.01% | 0.01% | -0.01% | 97.6% | 99.1% | 100.0% | 100.0% |
Class D | 0.01% | -0.24% | -0.31% | 97.2% | 99.6% | 100.1% | 100.0% |
Class F | #VALUE! | #VALUE! | #VALUE! | #NUM! | #NUM! | #NUM! | #NUM! |
Class TGM | #VALUE! | #VALUE! | #VALUE! | #NUM! | #NUM! | #NUM! | #NUM! |
The proposed method would introduce block shape dependent filtering scheme for HOG-based gradient computation.
It was commented that the introduced Sharr filter needs the multiplication operation whereas the current Sobel filter can be implemented without any multiplications. The proponent suggested that the coefficients of +/-10 and +/-3 can also be implemented as two additions rather than one multiplication.
It was reported by the proponent that the proposed filter can only bring gains for rectangular shaped blocks (i.e, does not work on square block shapes).
For small blocks (when 2x2 kernel is applied), filter is not changed in this proposal.
It was commented that the gains are quite small, with some small increase in implementation complexity (e.g. more filter coefficients and block shape dependency).
Further study was recommended.