JVET-X0135 Non-EE2: Adaptive intra MTS [B. Ray, V. Seregin, M. Karczewicz (Qualcomm)]
This contribution proposes to use variable number of MTS candidates for intra coded blocks depending on the position of last significant coefficient. Simulation results on top of ECM-2.0 are:
- AI (Y/U/V): -0.10%/-0.06%/-0.07%, EncT 100%, DecT 99%
- RA (Y/U/V): -0.03%/-0.02%/0.00%, EncT 95%, DecT 99%
- LDB (Y/U/V): -0.10%/0.13%/-0.05%, EncT 101%, DecT 98%
Current ECM uses only DCT2 when only DC coefficient is non-zero, otherwise all options. The proposal is about using less MTS candidates when the number of other (AC) non-zero coefficients depending on last coeff. Position (1, 4 or 6 are used). Reduction in run time for RA is achieved by some additional encoder optimization possible by this.
Have other configurations of variable number been tested? Yes, some other options, not extensively, was the best tradeoff found.
Would the same encoder optimization be possible in ECM anchor? Yes, but would give loss without the change.
Is it simple to extract the last coeff. position? It is parsed independent and ahead of the MTS signalling.
Transform kernels are the same as currently in MTS of ECM.
It was agreed to study this in an EE, also investigate encoder-only option, dependency on the thresholds of number of candidates, and possibly dependency of threshold on QP (e.g. signalling it).