An AI Experiment at an AI Conference
BBC R&D’s AI in Production project aims to understand how AI and Machine Learning will impact the broadcast industry. We’re trying to understand how machines will take on the simpler, more repetitive tasks in future production teams, freeing up humans to focus more on the creative aspects of their work. For example, perhaps AI-controlled cameras could be used to create basic shots, like simple close-ups of people moving around a stage, freeing up human camera operators to offer more elaborate options such as jib and dolly shots. Such a system would increase production values within a fixed budget, or broaden the range of output that broadcasters could offer.
A key question that needs to be answered before we can implement this kind of automation is, how do we teach computers what is interesting? Of all the things in a scene or on a stage, which will capture a viewer’s attention? One intriguing possibility is that when a show has a live audience, we might be able to get that audience themselves to help – by trying to identify what captures their attention. Understanding that could give us insights into the things that a viewer at home might want a television director to show to them.
At the moment, this is just an interesting idea, but the BBC, LFCI and Alan Turing Institute’s conference on AI, Society and the Media gives us a rare opportunity to carry out an experiment to test this concept. The organisers have kindly agreed to let us set up some cameras in the auditorium. We’ll use them to film part of the audience (who will have been informed about the experiment), and we intend to analyse those recordings after the event. In this analysis we’ll see if there are any patterns in the way that audience members direct their attention that could be useful in terms of giving AI algorithms a hint regarding what might be the most interesting parts of the stage, at any given moment.
There are a number of caveats here. We’ll need to judge the lighting levels in the venue carefully, so that we have enough light for our filming without interfering with the proper ambiance for the conference. We may struggle to get enough people in focus, or at high enough resolution for the results to be useful. It may even turn out that audience attention just isn’t that useful as a way of predicting what viewers would like to see! That’s part and parcel of doing scientific research, though – we’ll learn something useful, even if it’s just that our original idea was a bad one.
