Image flows: Machine Learning and Cinematic rhythm

This paper addresses the analysis of cinematic rhythm through deep neural networks. It introduces a system designed to be used by found-footage video artists as well cinema scholars who wish to describe and visualize rhythmic flows. Although arguably essential to the appreciation of cinematic art, the concept of rhythm is difficult to characterize. The system to be presented here enables the comparison and contrast of moving image sequences with respect to their rhythmic flow and so demonstrates how machine learning affords new ways of describing phenomena, such as visual rhythm, which would otherwise resist theoretical description. The presentation will describe the system's technical architecture, a hybrid convolutional-transformer network, and give examples of its use in media art.