Back to Search Document details
32nd Meeting: Hannover, DE, October 2023 2023-10-06 18:55
AHG8/AHG9: Truncated bit depth support SEI messages
Authors: D. Ding Tencent Xin Zhao Tencent Shan Liu Tencent G. Teniou Tencent S. Wenger Tencent Arianne Hinds
Abstract
Currently, video content is typically represented with eight or higher bit depth, such as 8-bit or 10-bit, which has been widely adopted for video communication perceived by human vision. However, it is observed through experimental results that, for machine vision/tasks, such as object tracking and detection, the quality score, e.g., measured by detection rate, shows lower sensitivity to the bit depth. Therefore, a lower sample value range, which can be represented by fewer bits per sample than 8 bits, can achieve a better trade-off between bitrate and machine task score.
Meeting Notes
JVET-AF0138 AHG8/AHG9: Truncated bit depth support SEI messages [D. Ding, X. Zhao, S. Liu, G. Teniou, S. Wenger (Tencent)]
JVET-AF0138 AHG8/AHG9: Truncated bit depth support SEI messages [D. Ding, X. Zhao, S. Liu, G. Teniou, S. Wenger (Tencent)]

Currently, video content is typically represented with eight or higher bit depth, such as 8-bit or 10-bit, which has been widely adopted for video communication perceived by human vision. However, it is observed through experimental results that, for machine vision/tasks, such as object tracking and detection, the quality score, e.g., measured by detection rate, shows lower sensitivity to the bit depth. Therefore, a lower sample value range, which can be represented by fewer bits per sample than 8 bits, can achieve a better trade-off between bitrate and machine task score.

In this contribution, it is proposed to include SEI messages signalling the actual sample value range (or bit depth) of the input video, which is represented by 8-bit or higher bit depth.

The idea is that only the n LSBs are used when an n-bit video is input into the decoder. It is verbally reported that this provides benefit in terms of improving BD-mAP.

From the experience with IBDI, such an approach would generally not have a benefit for video that is encoded for human consumption.

It appears that the benefit of this proposal would be mainly for VCM. It was negotiated in joint meetings with WG 4 that JVET would not define metadata that are specifically useful only for machine vision applications. This would be a typical case of an SEI message that should be managed within WG 4.

ARCHIVE :
._JVET-AF0138_v2.docx JVET-AF0138_v2.docx
Decisions
JVET-AF0138 AHG8/AHG9: Truncated bit depth support SEI messages [D. Ding, X. Zhao, S. Liu, G. Teniou, S. Wenger (Tencent)]
It appears that the benefit of this proposal would be mainly for VCM. It was negotiated in joint meetings with WG 4 that JVET would not define metadata that are specifically useful only for machine vision applications. This would be a typical case of an SEI message that should be managed within WG 4.
Citation