JVET-AO0071 AHG12: Simplification on TALF [L. Xu, N. Song, Y. Yu, H. Yu, D. Wang (OPPO)]
This contribution proposes to modify the filtering process of TALF to generate the offset without dependency on the output sample of SAO. The number of filter coefficients is reduced by 1. The proposed method is implemented on top of the ECM-19.0 software. The simulation results are summarized as below:
RA: {0.00%, -0.06%, -0.06%, 100%, 100%}
LDB: {-0.04%, 0.22%, 0.11%, 100.1%, 100.0%}
Class-by-class results are as follows:
Random Access Main 10 | |||||||
Over ECM-19.0 | |||||||
Y | U | V | EncT | DecT | EncVmPeak | DecVmPeak | |
Class A1 | 0.01% | -0.11% | -0.07% | 99.9% | 99.5% | 100.0% | 100.4% |
Class A2 | -0.01% | -0.07% | 0.00% | 100.2% | 100.9% | 100.1% | 100.5% |
Class B | 0.01% | -0.02% | -0.06% | 100.0% | 99.9% | 100.0% | 100.0% |
Class C | -0.01% | -0.06% | -0.08% | 99.9% | 99.7% | 100.0% | 100.0% |
Class E |
|
|
|
|
| ||
Overall | 0.00% | -0.06% | -0.06% | 100.0% | 100.0% | 100.0% | 100.2% |
Class D | 0.01% | -0.07% | -0.16% | 99.9% | 100.0% | 100.0% | 100.0% |
Class F | 0.11% | -0.20% | -0.19% | 99.6% | 99.1% | 99.8% | 100.1% |
Low delay B Main 10 | |||||||
Over ECM-19.0 | |||||||
Y | U | V | EncT | DecT | EncVmPeak | DecVmPeak | |
Class A1 |
|
|
|
|
|
|
|
Class A2 |
|
|
|
|
| ||
Class B | -0.08% | 0.02% | -0.34% | 100.3% | 99.9% | 100.1% | 100.0% |
Class C | -0.01% | 0.25% | 0.00% | 99.8% | 100.3% | 100.0% | 100.0% |
Class E | -0.03% | 0.51% | 1.00% | 100.2% | 99.9% | 100.2% | 100.0% |
Overall | -0.04% | 0.22% | 0.11% | 100.1% | 100.0% | 100.1% | 100.0% |
Class D | -0.01% | -0.32% | 0.04% | 101.0% | 99.6% | 99.9% | 100.0% |
Class F | 0.00% | -0.57% | -0.06% | 100.2% | 99.8% | 100.1% | 99.8% |
Current ECM-19.0 flow is as follows:
Dependency on SAO output is removed using the following proposed flow, also as a consequence filter tap is reduced by 1:
The proposed method would allow TALF to be performed in parallel to deblocking and BIF/SAO/CCSAO, improving parallelization of the loop filters in ECM.
It was commented that although the gain is relatively small, class-by-class gains are consistent, and this seems to be a straightforward simplification.
Cross checker reported matched results.
Potential reasons for the luma loss in class F and chroma loss in LDB configuration were unknown.
It was agreed to investigate this in an EE.