Abstract

Temporal Action Segmentation (TAS) assigns an action label to each frame in a sequence. While TAS has been extensively studied in contexts such as daily activities, sports, and surgical videos, its application to continuous dance motion derived from video-based skeleton data is still limited. This paper evaluates vision-based TAS for continuous dance motion using video-derived skeleton sequences, eliminating the need for wearable sensors. A labeled dance dataset was created from video clips, including multi-person scenes where each performer was extracted and treated as an individual sequence. The dataset also incorporates duet performances, which introduce greater variation in execution and recording conditions.To examine how pose representation and temporal modeling affect segmentation quality, we evaluated two skeleton settings (BODY-25 and COCO-18) and two representative Temporal Action Segmentation (TAS) models [Multi-Stage Temporal Convolutional Network++ (MS-TCN++) and Frame-Action Cross-Attention Temporal Modeling (FACT)] under a unified protocol. Both models effectively segmented dance actions from skeleton data. FACT consistently achieved higher segment-level performance than MS-TCN++, particularly in F1 and Edit scores, while BODY-25 generally outperformed COCO-18 across both models. Using BODY-25, FACT achieved F1@10 and F1@25 scores of 94.6 and an F1@50 score of 90.6, compared to MS-TCN++'s scores of 89.7, 89.7, and 89.1, respectively. These results demonstrate that both pose representation and temporal modeling materially affect segmentation quality in continuous dance motion.