Alexei (Alyosha) Efros and Frederick Alexander Hall and Maneesh Agrawala

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2018-102

August 7, 2018

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-102.pdf

Film historians and filmmakers study the visual style of past films to answer research questions or gain inspiration for new projects. To help such film professionals conduct large-scale analyses of visual style in films, we present Film Grok, a computational tool for labeling and analyzing narrative films. We automatically label a dataset of 620 films with key features of visual style (e.g., character framing, shot sequences) derived from filmmaking texts. To study these features in the broader context of the film, we provide narrative features such as dialogue, emotional sentiment, genre, and director. For example, we use our tools to show that the rise of TV in the 1950’s correlates with character framings that are on average 5% closer to the center of the screen and nearly 200% closer to the actor than they were in the 1930’s and 40’s. We show in another example that Westerns tend to use Extreme Long Shots at moments with 70% stronger negative sentiment than the rest of the film. Akira Kurosawa, a self-proclaimed student of American Westerns, furthers this trend, using Extreme Long Shots for moments with 400% stronger negative sentiment. We train an SVM to classify films based on genre and from this SVM extract the most discriminative shot sequences for each genre. Additionally, we use Film Grok’s labels to automatically produce supercuts and supergrids highlighting visual features of interest based on user queries.

Advisors: Maneesh Agrawala and Alexei (Alyosha) Efros

---

BibTeX citation:

@mastersthesis{Efros:EECS-2018-102,
    Author= {Efros, Alexei (Alyosha) and Hall, Frederick Alexander and Agrawala, Maneesh},
    Editor= {Pavel, Amy},
    Title= {A Tool for Computational Analysis of Narrative Film},
    School= {EECS Department, University of California, Berkeley},
    Year= {2018},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-102.html},
    Number= {UCB/EECS-2018-102},
    Abstract= {Film historians and filmmakers study the visual style of past films to answer research questions or gain inspiration for new projects. To help such film professionals conduct large-scale analyses of visual style in films, we present Film Grok, a computational tool for labeling and analyzing narrative films. We automatically label a dataset of 620 films with key features of visual style (e.g., character framing, shot sequences) derived from filmmaking texts. To study these features in the broader context of the film, we provide narrative features such as dialogue, emotional sentiment, genre, and director. For example, we use our tools to show that the rise of TV in the 1950’s correlates with character framings that are on average 5% closer to the center of the screen and nearly 200% closer to the actor than they were in the 1930’s and 40’s. We show in another example that Westerns tend to use Extreme Long Shots at moments with 70% stronger negative sentiment than the rest of the film. Akira Kurosawa, a self-proclaimed student of American Westerns, furthers this trend, using Extreme Long Shots for moments with 400% stronger negative sentiment. We train an SVM to classify films based on genre and from this SVM extract the most discriminative shot sequences for each genre. Additionally, we use Film Grok’s labels to automatically produce supercuts and supergrids highlighting visual features of interest based on user queries.},
}

EndNote citation:

%0 Thesis
%A Efros, Alexei (Alyosha) 
%A Hall, Frederick Alexander 
%A Agrawala, Maneesh 
%E Pavel, Amy 
%T A Tool for Computational Analysis of Narrative Film
%I EECS Department, University of California, Berkeley
%D 2018
%8 August 7
%@ UCB/EECS-2018-102
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-102.html
%F Efros:EECS-2018-102