JVET-AL0290 [AHG12] Combination of JVET-AL0174, JVET-AL0195 and JVET-AL0241 on Enhanced TIMD with NNIP and Optimized Block Vector Derivation [K. Naser, S. Puri, T. Dumas, M. Radosavljević, F. Le Léannec, E. François (InterDigital), J. Fu, Y. Zhao, J. Zhang, S. Ma (PKU), C. Huang (ZTE), D. Ruiz Coll, J.-K. Lee (Ofinno), Y.-H. Lin, C.-Y. Teng, K.-W. Liang, Y.-C. Yang (Sharp)] [late]
This contribution provides the combined results of the three contributions JVET-AL0174, JVET-AL0195, which are about TIMD improvement with NN-Intra and optimized block vector derivation.
The following results are obtained over ECM-16.1:
JVET-AL0195+ JVET-AL0241+ JVET-AL0174: AI: {-0.13%, -0.04%, -0.04%}
JVET-AL0195+ JVET-AL0241: AI: {-0.12%, -0.06%, -0.08%}
The following shows AI results in the individual contributions vs. the combination:
- Natural sequences
Contribution | Luma Performance | EncT | DecT |
JVET-AL0174 (TIMD+NNIP) | -0.02% | 100.3% | 100.6% |
JVET-AL0195 (TIMD+BV) | -0.08% | 101.1% | 101.0% |
JVET-AL0241(ITMP template type adaptation) | -0.04% | 100.2% | 100.1% |
-0.12% | 101.0% | 101.8% | |
JVET-AL0174+JVET-AL195+JVET-AL0241 | -0.13% | 100.4% | 100.5% |
- Class F and Class TGM averaged
Contribution | Luma Performance | EncT | DecT |
JVET-AL0174 (TIMD+NNIP) | N/A | N/A | N/A |
JVET-AL0195 (TIMD+BV) | -0.29% | 102.2% | 103.6% |
JVET-AL0241(ITMP template type adaptation) | -0.19% | 99.55% | 100.65% |
-0.48% | 98.9% | 100.0% | |
-0.47% | 97.6% | 98.6% |
It was asked if the runtime is reliable, as the combination shows reduction in enc/dec runtime whereas some individual tests show some runtime increase. According to the proponents, the combination runtime may not be reliable.
It was commented that, looking at the combination tests, the performance contribution from NNIP seems to be the least interesting.
For the BV-TIMD + NNIP combination, BV is applied at the template matching stage along with planar and DC, whereas NNIP is applied in the prediction stage when certain conditions are met (DC or planar is selected).
It was commented that detailed descriptions are not available in the proposals, and the proponents are requested to write clear EE descriptions on how tests are to be conducted in detail.
It was agreed to study this in an EE; within which the following tests should be conducted:
- NNIP to replace planar and DC in TIMD (from JVET-AL0174)
- BV replacing regular intra in TIMD (from JVET-AL0195)
- IntraTMP merge list improvement (from JVET-AL0195)
- Template adaptation for LIC + fusion for intraTMP mode (from JVET-AL0241)
- Combination of all of the above except the NNIP elementCombination of all of the above