Back to Search Document details
26th Meeting: by teleconference, April 2022 2022-04-28 11:03
EE1-related: Reduced complexity NN loop filter and ablation study
Authors: J. Ström Ericsson Du Liu Ericsson Mitra Damghanian Ericsson K. Andersson Ericsson Yun Li Ericsson P. Wennersten Ericsson Ruoyang Yu Ericsson
Abstract
This contribution presents an ablation study of the intra model in the neural network-based loop filter from JVET-Y0143. It is stated that the network is trained several times from scratch, each time changing a single variable, such as removing one of the inputs to the filter. Some changes are claimed to hurt BD-rate performance, such as removing the prediction input. It is further claimed that other changes do not seem to impact the BD-rate performance. As an example, the contribution states that removing the partitioning input lowers the complexity of the intra luma-filter without harming the BD-rate, giving BD-rate figures over the VTM-11.0 + newMCTF anchor of
References:
JVET-X2016 JVET-Y0143
Cited:
JVET-AA0023 JVET-AA0081 JVET-Z0011 JVET-Z0023 JVET-Z0203 X JVET-Z0234 JVET-Z2023
PATENTS:
AU2016244241C1 0.34 2020-12-17 AU2020269469B2 0.32 2026-01-29 ES2899581T3 0.30 2022-03-14 EP3650548B1 0.28 2022-03-09 US11389171B2 0.26 2022-07-19 AU2023201926B2 0.24 2025-08-14 US10131686B2 0.22 2018-11-20 KR102429859B1 0.20 2022-08-04
ARCHIVE v5 :
._.DS_Store ._JVET-Z0106_doubleDepth__0378_v3.xlsm ._JVET-Z0106_halfDepth__0378_v3.xlsm ._JVET-Z0106_noAttention__0378_v3.xlsm ._JVET-Z0106_noBS__0378_v3.xlsm ._JVET-Z0106_noPartitioning__0378_v3.xlsm ._JVET-Z0106_noPartitioning__0378_v5.xlsm ._JVET-Z0106_noPartitioning_noBS__0378_v3.xlsm ._JVET-Z0106_noPrediction__0378_v3.xlsm ._JVET-Z0106_one_attention__378_v3.xlsm ._JVET-Z0106_retraining__0155_v3.xlsm ._JVET-Z0106_retraining_longer_time_0378_v3.xlsm ._JVET-Z0106_v1.docx ._JVET-Z0106_v2.docx ._JVET-Z0106_v3.docx ._JVET-Z0106_v3_clean.docx ._JVET-Z0106_v5.docx ._JVET-Z0106_v5.pptx ._JVET-Z0106_v5_clean.docx .DS_Store JVET-Z0106_doubleDepth__0378_v3.xlsm JVET-Z0106_halfDepth__0378_v3.xlsm JVET-Z0106_noAttention__0378_v3.xlsm JVET-Z0106_noBS__0378_v3.xlsm JVET-Z0106_noPartitioning__0378_v3.xlsm JVET-Z0106_noPartitioning__0378_v5.xlsm JVET-Z0106_noPartitioning_noBS__0378_v3.xlsm JVET-Z0106_noPrediction__0378_v3.xlsm JVET-Z0106_one_attention__378_v3.xlsm JVET-Z0106_retraining__0155_v3.xlsm JVET-Z0106_retraining_longer_time_0378_v3.xlsm JVET-Z0106_v1.docx JVET-Z0106_v2.docx JVET-Z0106_v3.docx JVET-Z0106_v3_clean.docx JVET-Z0106_v5.docx JVET-Z0106_v5.pptx JVET-Z0106_v5_clean.docx
Citation