Hobby drones equipped with a camera are now a very popular tool around the world, while a group of researchers has developed an artificial intelligence-based framework that allows even more serious filmmakers to make shooting more efficient and cheaper. Today’s modern drones already have highly advanced capabilities to detect and track objects, even from the right distance. Because of this, they have become a very important tool for both amateur and professional filmmakers.
But controlling the camera of quadrocopters is often very cumbersome. Drones can walk different paths in the same scene, even if they carefully adjust their position, speed, and angle of their camera. This, in turn, can completely ruin the consistency of the recording. In search of a solution, researchers at Carniege Mellon University, the University of Sao Paolo and Facebook have worked together to develop a framework that can eliminate this problem. Using software simulations, a database was compiled from video clips with different recording types. Then, by involving a large group of people (crowdsourcing) and using artificial intelligence, a connection was established between predefined labels (“exciting”, “enjoyable”, “predictable”) and certain semantic descriptors. In the next step, the co-authors of the framework embarked on a series of experiments that included a special process, the determination of the minimum perceptually valid step size. This allowed the parameterization of the drone recordings. Then they built a data set of 200 video files and commissioned people recruited from Amazon Mechanical Turk to assign scores to the semantic descriptors.
The scores formed a machine learning model that linked descriptors to parameters that could control drones during recordings. Finally, the team installed the framework on a Parrot Bebop 2 drone and claimed to have achieved very good results with a variety of actors, activities and settings alike. Video production can be very expensive. In the United States, a short commercial can cost as much as $ 1,500 to $ 3,500, or about half a million to one million forints, and can be a serious burden for a small and medium-sized business. Artificial intelligence-based methods, on the other hand, can significantly reduce costs, so such solutions are likely to gain more ground in the future.