Intelligent highlights generation for soccer game

Abstract

Nowadays, with the evolution of technology, people have access to Internet everywhere. Thus, videos of soccer games have to be adapted to this media to improve customers’ ex- perience. Highlights are a particular way to provide short videos people can look at in the bus for example. These videos may be personalized to each user. The purpose is to create emotion to viewers. To do so, in this work, the methods and principles used by operators who manually generate highlights were studied. Moreover, statistics about the content people want to watch and about manual highlights has been extracted. From that basis, a list of the most interesting events in a soccer game, personalized for a particular user and for a given duration is established. Then, the selection of views is effectuated from the camera streams that are available in live production. Finally, these streams are turned into clips depending on sound and on a tool, called EmotionNet, created in this master thesis. EmotionNet is a convolutional neural network able to detect emotion sequences with more than 82% accuracy. This network is also used for the selection of contextual sequences which help viewers understand the context of the soccer game they are looking at as well as the story of the game. Finally, the results obtained with this algorithm were assessed. Evaluation is difficult in the context of this work because there is no objective ground truth: manual highlights depend on the feeling of the person who does them. The performances were quantitatively estimated task by task. They are satisfying, especially in the case of EmotionNet. The overall result is to be assessed subjectively by personal appreciation. The highlights generated are quite good, they present features of manual highlights and they are perzonalizable depending on the user.