Trailers, which are brief video clips that promote new films, are frequently essential components of film production firms' promotional tactics. A clip is said to be most effective when it summarises the movie's plot while also appealingly showcasing its artistic style and mood. Till now, it was a task undertaken by humans. However, computer scientists have started exploring how these promotional clips could be produced autonomously by the machines in recent times.
For making this happen, a research team has developed a model that uses an unsupervised, graph-based machine-learning algorithm to create short video clips. The new method is set to contribute significantly towards the Machine Learning As a Service Market. It would remove the unnecessarily burdensome task from the humans while also helping generate trailers that meet all the required criteria of a good video clip.
To find this method of automatic film trailer generation, the team divided the tasks into two parts. Firstly, they built a structure that could identify the movie's narrative style. Secondly, they taught it to predict the film's sentiment, i.e., its conveyed mood or feeling. The resultant technique thus was the product of both parts of the movie – videos and text extract taken from the movie's screenplay.
The method comprises two neural networks. One of them is responsible for processing multimodal shot representations taken from the movie's video stream. On the other hand, the second neural network analyses textual scene representations per the movie's screenplay.
The combination of two neural networks enables the identification of all the turning points of the movie. This includes the parts that are salient and are necessary to be included in the trailers. Generally, turning points in the film have an opportunity, a twist in the plan, a moment of critical decision, a significant challenge and a climax. It is necessary to properly merge all these in a trailer in an appealing manner without disclosing the movie's whole plot, and the team has done exactly that.
The researchers added that they would like to emphasize methods that help predict complex emotions such as terror, grief, joy, and loathing in the coming future. For now, they considered the positive and negative emotions as an alternative to emotions as there were no in-domain labeled datasets.