Back to Search Document details
29th Meeting: by teleconference, DE, January 2023 2023-01-12 20:10
EE1-related: Improvement on EE1-1.7
Authors: Zhihuang Xie OPPO Yue Yu OPPO H. Yu OPPO D. Wang OPPO
Abstract
In EE1-1.7, two aspects were proposed to improve the coding performance of NNVC-3.0 filter set 1. First, it is proposed to feed collocated blocks from reference frames into the CNN-based in-loop filtering. Second, the input samples of CNN can be flipped and the output samples of CNN can be flipped back. This contribution proposes a slice-level model switching method to further improve the performance of filter set 1. Compared with EE1-1.7, the performance is reported as below:
Meeting Notes
JVET-AC0066 EE1-related: Improvement on EE1-1.7 [Z. Xie, Y. Yu, H. Yu, D. Wang (OPPO)]

In EE1-1.7, two aspects were proposed to improve the coding performance of NNVC-3.0 filter set #1. First, it is proposed to feed collocated blocks from reference frames into the CNN-based in-loop filtering. Second, the input samples of CNN can be flipped and the output samples of CNN can be flipped back. This contribution proposes a slice-level model switching method to further improve the performance of filter set #1. Compared with EE1-1.7.1, the performance is reported as below:

RA: Y 0.00%, U 0.00%, V 0.00%, EncT 100%, DecT 100%;

LDB: Y -0.93%, U 0.01%, V 0.06%, EncT 103%, DecT 100%.

For RA, the method is not used.

The flipping is already used in EE1-1.7

The “normative” change would be a slice-level flag for signalling the model, and at the decoder the POC needs to be checked. This would be some kind of HLS control. Otherwise, the proposal is more about encoder optimization on top of EE1-1.7.

Several experts supported this as an interesting idea to improve the performance in LDB.

This could be a candidate for adoption provided that EE1-1.7 would be adopted. It was reported verbally during session 27 that the gains are also retained (or with partial results slightly higher) for LDB when the method is implemented on top of the combination 1.1.5/1.1.7 (as per JVET-AC0310). It was requested to provide these results in a new input (see JVET-AC0355).

Cited:
JVET-AC0304 X JVET-AC0308 X
ARCHIVE :
JVET-AC0066-presentation.pdf JVET-AC0066-v1.docx JVET-AC0066-v2.docx JVET-AC0066_vs_EE1-1.7.1.xlsm JVET-AC0066_vs_nnvc-3.0_anchor.xlsm JVET-AC0066_vs_nnvc-3.0_NnlfOption2.xlsm
Decisions
adopted
This could be a candidate for adoption provided that EE1-1.7 would be adopted. It was reported verbally during session 27 that the gains are also retained (or with partial results slightly higher) for LDB when the method is implemented on top of the combination 1.1.5/1.1.7 (as per JVET-AC0310). It was requested to provide these results in a new input (see JVET-AC0355).
Citation