Back to Search Document details
33rd Meeting: by teleconference, CH, January 2024 2024-01-19 23:32
[AHG11] Study on lower-complexity NNLF
Authors: Zhihuang Xie OPPO Yue Yu OPPO H. Yu OPPO D. Wang OPPO
Abstract
This contribution reports the experimental results of a lower-complexity NNLF, with computational complexity about 5.0 KMAC/pixel. The training and inference processes are identical to the NNVC-7.1 LOP.2. Compared with VTM anchor, the simulation results of the proposed lower-complexity NNLF are summarized as follows:
Meeting Notes
JVET-AG0057 [AHG11] Study on lower-complexity NNLF [Z. Xie, Y. Yu, H. Yu, D. Wang (OPPO)]

This contribution reports the experimental results of a lower-complexity NNLF, with computational complexity about 5.0 KMAC/pixel. The training and inference processes are identical to the NNVC-7.1 LOP.2. Compared with VTM anchor, the simulation results of the proposed lower-complexity NNLF are summarized as follows:

Compared with VTM-anchor:

AI: -3.32% (Y) , -4.19% (U), -3.42% (V), 81% (EncT), 1631% (DecT)

RA: -2.89% (Y), -3.83% (U), -2.29% (V), 104% (EncT), 3473% (DecT)

Encoder run time is likely unreliable.

It was agreed to study this in the EE along with JVET-AG0155 and JVET-AF0206). Integerization also to be studied. It was suggested that for ultra low operation, it would be interesting to also investigate more aggressive integerization, such as 8/10 bits (clarify if that is possible with SADL).

ARCHIVE :
JVET-AG0057(NnIntraEnable)_vs_VTM-anchor.xlsm JVET-AG0057-v2.docx JVET-AG0057.pptx JVET-AG0057_vs_VTM-anchor.xlsm
Decisions
It was agreed to study this in the EE along with JVET-AG0155 and JVET-AF0206). Integerization also to be studied. It was suggested that for ultra low operation, it would be interesting to also investigate more aggressive integerization, such as 8/10 bits (clarify if that is possible with SADL).
Citation