Please note: This PhD defence will take place in DC 2314.
Andreea Pocol, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Lesley Istead & Craig S. Kaplan
Despite extensive research affirming colour-emotion associations and the importance of colour in film, there is no mechanism for retrieving or recommending films based on their colour palette. A primary obstacle to such a mechanism is that palettes for videos tend to be dominated by statistically prevalent tones that are unmemorable to viewers, and ultimately unable to capture and convey the visual essence of the cinematic content they represent. Moreover, current colour palette generation methods are uninterpretable ‘black boxes’ to users, leaving users unable to trace how colours were extracted from the source.
This thesis introduces a novel method for colour palette generation from video, which has demonstrated greater accuracy than the industry standard—as validated by a user study—in extracting memorable colours from visually rich movie trailers. We work with movie trailers as high-density visual summaries of full-length movies, allowing us to process the trailers for∼73,220 colour films released before 2025. Palettes generated using our method have revealed interesting trends in film colour use across decades, directors, and genres. Our method has proven useful in the retrieval task of identifying the best colour-based match for a given trailer, which has applications in movie recommendation systems. Two user studies confirmed, with statistical significance, that colour-based movie recommendations are a desirable feature for users, and adding colour as a feature to movie recommendations is effective. We present a web application, RefinedPalette, that makes the palettes interpretable, showing the frames contributing to each colour swatch in a palette, and offers colour-based movie browsing and recommendations. RefinedPalette was evaluated in a user study through a USE (usefulness, satisfaction, and ease of use) questionnaire, and results were positive with statistical significance in all three categories.