fbpx

Code: EDI-2020-20-VRT_2

Domain: Internet & Media

Summary

Automatic metadata generation concerning the experts and interviewees that are part of our daily news bulletin as well as how long they were present on screen.

Proposed by

 

VRT produces a large amount of content on a daily basis, ranging from radio content to news articles to video content. One of the main challenges related to this is making that content easily retrievable, both for the media user and for our editing teams. Typically, metadata is used to achieve this, for example program titles, content genre and other types of tags. Specifically for video, our news team is interested in the automatic metadata generation concerning the experts and interviewees that are part of our daily news bulletin as well as how long they were present on screen.

Description

VRT produces a large amount of content on a daily basis, ranging from radio content to news articles to video content. One of the main challenges related to this is making that content easily retrievable, both for the media user and for our editing teams. Typically, metadata is used to achieve this, for example program titles, content genre and other types of tags. Specifically for video, our news team is interested in the automatic metadata generation concerning the experts and interviewees that are part of our daily news bulletin as well as how long they were present on screen.

Data

The challenge has the following sample datasets available for download

Data: initially, 31 videos of our evening news bulletin with time coded closed captioning in Dutch. This number of videos will be increased later on.

Expected outcomes

Goal of the challenge: automatic recognition of the people that are present in our news bulletin. This concerns both the gender of the hosts, experts and interviewees present in the videos, as well as their name and function. Since a lot of these people are not world famous, it poses some difficulty to attach a name to their face. The solution should then also provide an easy way to name new interviewees and experts so they can automatically be recognized in the future if they should return. This could be based on visual information that typically identifies the interviewee. Secondly, we do not want to automatically recognize all people present on screen, but only the people that are talking during the news bulletin and not the people in the background during the interview for example. Ultimately, a dashboard should be provided in which we can see how many minutes a man or a woman was present (talking) on screen for one or multiple recordings and queries concerning how long a specific individual was on screen should also be possible.

How do we apply?

Follow by Email
Facebook
Twitter
YouTube
SOCIALICON