"I Want to Dance": Comparative K-Pop Choreography Analysis Through Deep-Learning Pose Estimation

Critics have long noted the strong visual aspects of K-pop, with the videos for newly released songs garnering millions of hits in a very short time span. A key feature of many Kpop videos is the dancing. Although many of the official videos are not solely dance focused, incorporating aspects of visual storytelling, nearly all of Kpop videos include some form of dance. In addition to the "main" video for a Kpop release, the release of a dance video, or a dance rehearsal video, focusing exclusively on the dances has become common practice. These videos allow fans to learn and practice the dance, thereby increasing the kinesthetic connection between fans and their idols.

At the same time, it affords an opportunity to explore the "dance vocabulary" of Kpop dances. While there are well-known Kpop choreographers who work with the Kpop idols to create their dances, there is little documentation of these dances beyond the dance videos themselves. In our work, we develop a series of methods for (a) identifying dance sequences in Kpop videos, irrespective of whether they are dance videos (b) develop a series of classifiers for the navigation of a large scale Kpop video corpus and (c) apply deep learning methods to identify dancers and their body positions.

Taken together, these approaches pave the way for the development of a macroscope for the study of Kpop videos, allowing researchers to identify patterns in the Kpop space, explore dynamic change in features such as color space, or interrogate the differences in visual representations of male and female performers at an aggregate scale. Importantly, as pose estimation has become more accurate, these methods allow us to begin the process of inferring the dance vocabulary of Kpop and start the process of tracing transcultural choreographic flows.

• Timothy R. Tangherlini (University of California, Berkeley)
• Peter Broadwell (Stanford University)

Vidéo MemoRekall