MIT-Skate

Early benchmark by Pirsiavash et al. (2014) with fine-grained action labels for singles programs.

Tasks
Action recognition, Element segmentation, Temporal detection
Modalities
Broadcast RGB video
Size
150 labeled programs (~18 hours)
Coverage
ISU Grand Prix Series 2010–2012
Availability
request
License
Research only (contact authors)
Maintainer
MIT Computer Science & Artificial Intelligence Laboratory
Released
Mar 2014

Highlights

Notes for Practitioners

The dataset ships on request with MPEG-4 clips capped at 720p; expect visible compression artefacts. Label timing is coarse (±0.5s), so modern detection pipelines may require manual refinement. Cite both the dataset and the associated CVPR workshop report when publishing results to honour the original distribution terms.