Online-Event / June 09 – 10, 2020
The MDN Workshop is the annual meeting point for developers working on Metadata and Artificial Intelligence in broadcasting. It is open to the public.
The event is organized by the EBU Strategic Programme on Media Information Management and AI (MIM-AI) and the Metadata Developer Network (MDN), an active community for developers to share knowledge, learn from their peers, get feedback and collaborate on metadata-related projects.
Fraunhofer IDMT's experts participate with two presentations:
06/09/2020, 13:30 – 14:00
The Pitfalls of AI for Automatic Metadata Extraction
There is no doubt on the potential of AI for automatic metadata extraction (AME) in media assets. AI can help to recognize humans in image, video or audio; to analyse the spoken word; to auto-tag objects, events or semantical properties, just to name a few. However, while applying the out-of-the-shelf solutions for AME, customers often experience unexpected or undesired behaviour and are not satisfied with the results. This mismatch between the expectations and results has roots in various types of bias that are introduced along the AI pipeline. In the presentation, I would like to discuss the most significant types of bias for black-box data-driven AI pipeline, explain why those might cause AI models to fail and show the ways to cope with them.
06/10/2020, 14:30 – 15:00
Audio Partial Matching (and other Audio Tools) for Program Analysis
The presentation will show how audio partial matching tools can be used for various purposes, including de-duplication, rights tracking and reuse detection. More specifically, it will describe proof-of-concepts for two specific use cases which have been realized and tested recently:
(1) detection and localization of original program and cleanfeed segments within rundowns, which was tested e.g. using material from ORF, and, (2) analysis of radio and TV programs with respect to repetitions and structure, which was tested using material from German radio stations.