As the consumption of video contents rises and consumers’ behaviour changes, video broadcasters must find new strategies to engage with their viewers. In this context, metadata are used to provide customers with personalised, relevant contents. 

More recently, content-descriptive metadata have enhanced the personalisation of services and have offered providers new monetisation opportunities thanks to, e.g., more effective content recommendations and contextual advertising. So far, descriptive metadata have been applied to pre-recorded contents, while they still represent a huge amount of unexploited valuable information for live streams. 

This paper highlights how deep learning technologies can add value to information contained in each video frame of both live and non-live contents, automatically providing detailed descriptive metadata. This technology opens up the opportunity for broadcasters to offer new services and an innovative way of video consumption. 


Over the last years, the access to video contents has rapidly grown and customer behaviour in consuming these contents has changed accordingly. To give an idea of the growth of digital media, according to Cisco, by 2021, 82% of global Internet consumption will be video content (Cisco). PwC’s reports that the total worldwide subscription spending on Netflix and other over-the-top (OTT) subscription video-on-demand (SVOD) services grew by 33.8% in 2014 and 32.3% in 2015, that is 77% in two years (PwC).

One of the challenges broadcasters must face is how to use the benefits offered by these new models to engage with their viewers and generate new revenue opportunities from the expanding customer base.

Metadata are the key to achieve these goals. Content attributes together with viewer information are currently used to offer a higher quality viewing experience, providing customers with personalised, relevant contents. More recently, metadata describing the video content itself have been used to enrich video contents by providing external links and relevant, detailed information.

Furthermore, content-descriptive metadata offer broadcasters new monetisation opportunities, enabling more effective recommendations, contextual advertising, and new content-management tools.

So far, however, descriptive metadata have been used for VOD and non-live contents, since current metadata extraction from video relies heavily on human tagging and annotation; even if this procedure is highly precise, it can be unfeasible for large volumes of content. On the other hand, descriptive metadata still represent a huge amount of unexploited valuable information for live stream, which needs a flexible and efficient feature-extraction method able to react in real-time.

Download the full technical paper below