Back to Search Document details
43rd Meeting: Geneva, July 2026
AHG9: Classification label info SEI
Abstract
Two options are proposed to address the use case described in JVET-AP0085 “AHG9: Subject Recognition Information (SRI) SEI message.” In the first option, a classification label information SEI message is proposed to describe the subject or scene of a picture using a classification label. In this option, label types corresponding to each of several popular image classification datasets are specified, and applications may indicate custom classification label types by signaling a URL with a link to a CSV file containing a label list. In the second option, the text descriptions SEI message is extended to add an object label purpose.
JVET-AQ0149 AHG9: Classification label info SEI [J. Boyce, M. M. Hannuksela, T. Biatek (Nokia)]

Chaired by S. Deshpande during 16:30-17:00 on 12 July 2026.

Two options are proposed to address the use case described in JVET-AP0085 “AHG9: Subject Recognition Information (SRI) SEI message.” In the first option, a classification label information SEI message is proposed to describe the subject or scene of a picture using a classification label. In this option, label types corresponding to each of several popular image classification datasets are specified, and applications may indicate custom classification label types by signaling a URL with a link to a CSV file containing a label list. In the second option, the text descriptions SEI message is extended to add an object label purpose.

Since normative referencing of these dataset defined label would be problematics, actual values may be define in our standard, or an application defined list via URL to a csv file may be signaled. Integer or string label can be assigned. Only one label type per SEI is proposed.

For option 2, the text description is not free form string, but further extends the TDI SEI.

It was asked if the lists were created for images or for video. That seems to be the case.

How will decoder use this information: There may be specific image quality mechanisms which are done based on this.

If such information is useful, will it be associated with regions or entire picture: this is the main subject or scene.

Can annotated region SEI be used or extended instead? It is possible.

Is there any licensing issue in using the labels from these datasets. Also they could be added as Bibliography instead of normative references. Also it was commented that this classification may be tied to version of that dataset and it could be fast changing/ moving.

If you don’t have a standard use of the label how could the manufacturer use it.

There was a concern that why these 3 datasets are chosen. There are many others.

Extending TDI with additional syntax elements in an extension part for this functionality was not favored.

It was requested to address the above questions in a further study.

References:
PATENTS:
WO2026087151A1 1.00 2026-04-30 WO2026087636A1 1.00 2026-04-30 US20260012645A1 0.34 2026-01-08 WO2025219777A1 0.32 2025-10-23 WO2026046574A1 0.28 2026-03-05 WO2025215511A1 0.24 2025-10-16 US20260012647A1 0.22 2026-01-08 WO2026082343A1 0.20 2026-04-23
Decisions
It was requested to address the above questions in a further study.
Citation