Back to Search Document details
43rd Meeting: Geneva, July 2026
Cross-check of JVET-AQ0182 (EE1-4.1: adaptive quantization and hardware optimization for NNIP)
Authors: Aasaipriya Chandran Huawei
Abstract
This contribution reports the results of the cross-check of JVET-AQ0182 [1]. In NNVC-17.1 and ECM-20.0, in the Neural Network-based Intra Prediction (NNIP) implementation, layer activations use 16-bit signed integer, matrices of weights use 16-bit signed integer, and accumulators inside vector-matrix multiplications use 32-bit signed integer. In this design, accumulator overflow may occur. Moreover, high hardware area is required. JVET-AQ0182 [1] guarantees no accumulator overflow and reduces the hardware impact in each of its proposed 5 tests, namely EE1-4.1.1, EE1-4.1.2, EE1-4.1.4a, EE1-4.1.4b, and EE1-4.1.4c.
Meeting Notes
JVET-AQ0218 Cross-check of JVET-AQ0182 (EE1-4.1: adaptive quantization and hardware optimization for NNIP) [A. Chandran (Huawei)] [late]

EE1 related and beyond-EE contributions: Neural network-based video coding (6)

Contributions in this area were discussed during 1120–1225 on Wednesday 8 July 2026 (chaired by July).

References:
JVET-AF2016 JVET-AQ0182
Cited:
JVET-AQ0011 JVET-AQ0023
PATENTS:
CN117574976B 0.34 2024-04-30 CN112840358B 0.32 2023-05-23 US10469854B2 0.30 2019-11-05 EP4490701A1 0.28 2025-01-15 US11803734B2 0.26 2023-10-31 RU2759218C2 0.24 2021-11-11 CN109643443B 0.22 2024-03-08 CN119031147B 0.20 2025-05-16
ARCHIVE v2 :
Analysis_okto_1920x1080_10bits_420_BT709_chroma_left.xlsx Exp1_EE1-4.1.1_DynamicA10W8.xlsm Exp2_EE1-4.1.2_DynamicA8W8.xlsm Exp3_EE1-4.1.4a_StaticA10W8.xlsm Exp4_EE1-4.1.4b_StaticA12W8.xlsm Exp5_EE1-4.1.4c_StaticA14W8.xlsm JVET-AQ0218-v2.docx Result_okto_Anch_NNVC17.1-NNCfg_vs_EE1-4.1experiments_RASheet_f1.xlsm Result_okto_Anch_NNVC17.1-NNCfg_vs_VtmCfg-NNIP0-NNLF0_RASheet_f1.xlsm
Decisions
Contributions in this area were discussed during 1120–1225 on Wednesday 8 July 2026 (chaired by July).
Citation