Trailer and short video clips that anticipate the release of a film are a crucial activity in the promotion strategy of a film company. Until now they have remained a product of human activity, which is subject to competition from machines in many areas. But researchers at the University of Edinburgh have developed a model based on a pair of artificial neural networks that can automatically build cinematic trailers. In order to achieve their goal, Pinelopi Papalampidi, Frank Keller and Mirella Lapata claim to have been inspired by the creative process of the editors that identifies the most important parts of the narrative structure, namely the key events, and the process connected to the flow of emotions provoked. from the different scenes of a film.
The role of neural networks is to analyze, on the one hand, the visual representation of the various frames of a film, and on the other hand the textual representation based on the script. The result is the production of trailers that are able to identify the salient and most interesting aspects of the work to be summarized in a promo thanks to the action of an unsupervised automatic learning algorithm. Which, in a nutshell, means a computer capable of learning complex processes and patterns without constant human guidance.
The researchers subjected the fruit of their activity to a series of tests in order to evaluate the effectiveness of the technique adopted. In all, trailers for 41 different films were made.
In their work, described by a document published on arXiv, they show an example of how they have broken down Stanley Kubrik’s thriller masterpiece “The Shining” into various sequences and graphs in order to obtain a presentation clip. The researchers did not include in their program a selection criterion to completely exclude any spoiler scenes in order to avoid revealing essential contents of the plot to the viewer in advance. The trailers, however, were reviewed by juries of people selected through the Amazon Mechanical Turk service by comparing them with other videos generated by supervised learning algorithms. The good feedback received from human judgment supports a work that can be made available to film companies, especially independent cinema, to produce cheap promos.
For the production of higher quality automatic trailers, the researchers argue, the technique will have to be refined by improving the model in various aspects, and in particular the ability to recognize emotions such as pain, terror and joy more distinctly.
.