Back to Search Document details
21st Meeting: by teleconference, January 2021 2021-01-07 09:38
Balancing Luma-Chroma Channel Coding for DNN-based Intra-Frame Coding in JVET-U0079
Abstract
This contribution document presents a reformulation of the loss function used to train end-to-end DNN for YUV 4:2:0 intra-frame compression. It is asserted that by adjusting the weight between luma and chroma distortion (in mean-square error), a more balanced RD trade-off can be achieved.
JVET-U0080 Balancing Luma-Chroma Channel Coding for DNN-based Intra-Frame Coding in JVET-U0079 [A. K. Singh, H. E. Egilmez, M. Coban, M. Karczewicz (Qualcomm)]

This contribution document presents a reformulation of the loss function used to train and end-to-end DNN for YUV 4:2:0 intra-frame compression. It is asserted that by adjusting the weight between luma and chroma distortion (in mean-square error), a more balanced RD trade-off can be achieved.

Based on JVET-U0079 method 1, this contribution experimented with different weights for luma and chroma in loss function. In U0079, a ratio of Y:U:V weights of 4:1:1 was used, and this contribution changes the ratio to 2:1:1.

It was reported that, with this 2:1:1 ratio, (Y, U, V) performance over HM-16.20 becomes (−9.53%, 9.97%, −13.60%), compared to (−12.41%, 46.07%, 13.65%) as reported for JVET-U0079 method 1.

ParkRunning and Campfire chroma performance was also significantly improved, though chroma performance still shows a loss.

It was commented that non-equal weights between U and U components may help to reduce the imbalance as well.

Further study of YUV 4:2:0 coding in the context of the E2E framework is encouraged.

Decisions
Further study of YUV 4:2:0 coding in the context of the E2E framework is encouraged.
Citation