JVET-X0120 AHG12: On sign prediction [M. G. Sarwer, Y. Ye, J. Chen, R.-L. Liao (Alibaba)]
This contribution proposes a sorting-based approach to identify the signs to be predicted. In the proposed method, after decoding absolute values of the levels, the transform block (TB) is sorted based on the qIdx values and the first maxNumPredSigns signs of the sorted TB are predicted and rest of the signs are EP coded. This contribution also extended the sign prediction area of a TB from 4x4 block to maximum 32x32 block.
Following results are reported:
- AI: -0.11% (Y), -0.10% (U), -0.08% (V), 105% (EncTime), 105% (DecTime)
- RA: -0.14% (Y), -0.05 % (U), -0.05 % (V), 102% (EncTime), 107% (DecTime)
- LB: -0.16% (Y), 0.16 % (U), 0.09 % (V), 102% (EncTime), 106% (DecTime)
Current sign prediction in the ECM gives approx. 0.8%, which increases encoder run time by approx. 9% relative to the VTM. The new proposal has a worse tradeoff of complexity vs. compression benefit.
It was agreed to study this in an EE along with JVET-X0150; also in terms of the complexity tradeoff