Back to Search Document details
39th Meeting: Daejeon, KR, March 2025 2025-03-28 09:15
Non-EE2: TIMD fusion with neural network based intra prediction
Authors: Yi-Hao Lin Sharp Kai-Wen Liang Sharp Chih-Yu Teng Sharp Yu-Chiao Yang Sharp Karam Naser InterDigital T. Dumas InterDigital Edouard François InterDigital F. Le Léannec InterDigital
Abstract
This contribution proposes to replace the non-angular predictor (Planar or DC) with the neural network based intra prediction (NNIP) predictor in the template-based intra-mode derivation (TIMD) fusion process.
Meeting Notes
JVET-AL0174 Non-EE2: TIMD fusion with neural network based intra prediction [Y.-H Lin, K.-W Liang, C.-Y Teng, Y.-C Yang (Sharp), K. Naser, T. Dumas, E. François, F. Le Léannec (InterDigital)]

This contribution proposes to replace the non-angular predictor (Planar or DC) with the neural network based intra prediction (NNIP) predictor in the template-based intra-mode derivation (TIMD) fusion process.

On top of ECM-16.1, simulation results of the proposed method areAI : -0.02% 0.03% 0.01%, EncT: 100.3%, DecT: 100.6%.

The proposal replaces (instead of as an addition) Planar or DC with NNIP. This replacement is only applied to default TIMD, but not to TIMD merge and TIMD SAD.

It was commented that the increase in runtime is likely due to NNIP being more complex than planar and DC prediction.

This contribution is related to JVET-AL0195, JVET-AL0241, and JVET-AL0290, with the last contribution being the combination of this proposal with JVET-AL0195 and JVET-AL0241. See the notes under JVET-AL0290.

Cited:
JVET-AL0290
ARCHIVE :
JVET-AL0174-v3.docx JVET-AL0174.xlsm
Decisions
This contribution is related to JVET-AL0195, JVET-AL0241, and JVET-AL0290, with the last contribution being the combination of this proposal with JVET-AL0195 and JVET-AL0241. See the notes under JVET-AL0290.
Citation