BACKGROUND: Automatic recognition of skill movements in artistic swimming has become increasingly important after the 2023 rule revision, in which difficulty scores are multiplied by execution only when movements are correctly identified as valid. However, the number of qualified difficulty technical controllers is limited, and their workload is substantial. This study aimed to design and construct a fine-grained video dataset, defined as a dataset containing frame-level temporal annotations of individual skill movements and rule-defined action codes, to support future AI-based recognition of skill movements in artistic swimming. METHODS: Competition videos were recorded at Full HD resolution and 60 fps from a viewpoint equivalent to that of a technical controller. Videos from international and national competitions were collected, resulting in 481 routines across solo, duet, team, and acrobatic events. Certified technical controllers annotated the start and end times of each skill movement and assigned rule-defined codes using a dedicated web-based annotation system with frame-level adjustment functions. RESULTS: The dataset achieved high scene counts and high code coverage for Hybrid elements. In contrast, acrobatic categories—particularly pair acrobatics—showed lower scene counts and limited code coverage, indicating class imbalance. CONCLUSIONS: The constructed dataset provides a structured and rule-consistent foundation for future AI-based recognition of skill movements in artistic swimming. However, additional data collection and validation through AI model development and testing will be necessary before practical implementation in judging environments
© Copyright 2026 XVth International Symposium on Biomechanics and Medicine in Swimming Proceedings. Alle Rechte vorbehalten.