Highlights
- Ten labeled jump categories covering the most common single to quad elements.
- Temporal boundaries for preparation, take-off, flight, and landing phases.
- Includes Grade of Execution (GOE) annotations aligned with ISU scoring guidelines and aggregated by panel.
Notes for Practitioners
Baseline classifiers for this dataset rely heavily on pose quality. Fine-tuning a figure skating–specific pose backbone (for example, MotionBERT or RTMPose trained on skating clips) usually improves jump type accuracy by 4–6 percentage points compared with generic COCO-pretrained models.
The maintainers share access on request to research groups willing to abide by an athlete privacy agreement. When requesting access, prepare a secure storage plan and outline how you will anonymise broadcast footage before public release.