The researchers applied their method to a dataset of ten different dance genres, including Breakdance, Popping, and Krump. They used the AIST++ dataset, which contains over 1,400 sequences of dance footage captured at 60 frames per second. It is worth noting that some genre labels reflect overlapping or loosely defined categories within dance communities.
Baker handcrafted interpretable, statistical features of full-body movement, grounded in dance expertise. These features fall into four main categories. The first is movement of the sacrum, which is the base of the spine. The second covers movement of the extremities, such as the wrists and ankles. The third category includes angular momentum around the sacrum, which captures how a dancer’s body twists relative to their sacrum. The fourth group covers how expanded the body is from the sacrum.
Baker and his team built a machine-learning classifier using these features to identify which genre a dance sequence belonged to. Their model achieved a remarkable 76% accuracy in distinguishing among the ten genres, where chance performance would be 10%.
To evaluate how well humans can distinguish genres from movement alone, the team also conducted an online study. Participants with experience watching or performing hip-hop dance were shown silent, stick-figure renditions of the same sequences used to test the model. On average, these human observers achieved 38% accuracy – half that of the computational model.
Using a technique to measure feature importance called Shapley Additive Explanation, the team found that expandedness – how far a dancer’s limbs extend from their center – was the most predictive of genre, followed by sacrum motion and variability in ankle height.
By reducing their 17 features to three dimensions using a mathematical technique called Singular Value Decomposition, Baker’s team also created a ‘similarity space’, where different genres could be visualized and compared. This could be used to explore how hip-hop dance has evolved and how it might develop in the future.
The team’s approach offers a new window into how dance genres differ from one another, and highlights how these genres have evolved not just through associated attire, music, or settings, but in the look and feel of the movements.
This new method offers an interpretable, low-cost approach to data-driven study of full-body movement, and could be applied beyond dance to other domains of human movement, potentially informing movement training and therapies for dancers, athletes, and even patients undergoing physical rehabilitation.
