JVET-V0084 CE-related: On Rice parameter selection for regular residual coding (RRC) at high bit depths [A. Browne, S. Keating, K. Sharman (Sony)]
This contribution describes modifications to the method for the selection of Rice parameters for regular residual coding, for consideration in VVC version 2. The modifications offer BD-rate gains when residual coefficient values become larger but do not affect version 1 operating points. It is suggested that the modification extends the operating point of the existing residual coding method to be applicable at all currently considered bit depths. A further modification is described which alters the template function used to calculate a clipped sum of neighbouring coefficients used during Rice parameter evaluation. Results are provided both with and without this further modification.
Using the high bit depth, high bit rate CTC the following AI results were obtained for the low QP range against the CE anchor:
With template modification:
- PQ: -0.29%/-0.18%/-0.18%
- HLG: -0.56%/-0.34%/0.35%
- SVT-12: -1.67%/-1.26%/-1.27%
- SVT-16: -2.94%/-2.09%/-2.07%
Without template modification
- PQ: -0.26%/-0.16%/-0.16%
- HLG: -0.53%/-0.33%/-0.34%
- SVT-12: -1.65%/-1.26%/-1.26%
- SVT-16: -2.94%/-2.10%/-2.08%
Based on CE-1.6, history based method using counters, but is claimed to have various simplifications. Operates on sub-TU basis. Improved results compared to CE-1.6. Compared to CE-1.4 (which is better than CE1.6 generally), it is sometimes better, sometimes worse.
Would require more analysis on complexity in comparison to other proposals and the CE anchor.
Why is decoder runtime increased significantly for SVT class? Analysis shows that more 4x4 blocks are used than with other methods, which is closer to the worst-case scenario.
No specific action was taken on this, as the method adopted from the CE fulfils the purpose.