Datasets:

OpenSportsLab
/

OSL-loc-FigureSkating

pretty_name: FigureSkating Precise Event Spotting
license: other
task_categories:
  - video-classification
tags:
  - figure-skating
  - precise-event-spotting
  - action-spotting
  - opensportslib
configs:
  - config_name: fs_comp
    data_files:
      - split: train
        path: metadata/fs_comp/train.parquet
      - split: validation
        path: metadata/fs_comp/valid.parquet
      - split: test
        path: metadata/fs_comp/test.parquet
  - config_name: fs_perf
    data_files:
      - split: train
        path: metadata/fs_perf/train.parquet
      - split: validation
        path: metadata/fs_perf/valid.parquet
      - split: test
        path: metadata/fs_perf/test.parquet

FigureSkating precise event spotting

The complete 371-performance benchmark over 11 original competition broadcasts, with both official protocols and four event classes: jump takeoff/landing and flying-spin takeoff/landing.

Protocol	Train	Validation	Test
FS-Comp	178	59	134
FS-Perf	183	94	94

The protocols partition the same performances differently; media is shared and stored only once. Do not combine their training and evaluation sets.

Media and timing

Original 1920x1080, 25-fps source videos are in media/. Historical VP9 format 248 is used where available. The two women's World Championship broadcasts from 2017 and 2018 use available H.264 format 137 at the same resolution and frame rate, from the same original YouTube IDs. This encoding difference is recorded; bit-identical RGB with historical VP9 input is not claimed.

Performance clips are represented by exact half-open source-frame intervals, avoiding another video re-encoding. OSL event position_ms is absolute on the source video; the loader rebases it through metadata.intervals to each logical performance. reference_frame retains the native clip-relative event index, and source_frame retains the absolute source frame. At 25 fps one frame is exactly 40 ms. Native annotations remain in annotations/fs_comp/ and annotations/fs_perf/.

Each source has verified dimensions, FPS, interval bounds and a SHA-256 hash. This does not constitute a full decode of every hour of source media. Exact preprocessing and the selected decoder remain part of the experiment recipe.

OpenSportsLib loading

Use the OpenCV backend for interval-aware localization. Current DALI localization does not support these logical intervals.

from opensportslib.tools import download_dataset_split_from_hf
download_dataset_split_from_hf(
    repo_id="OpenSportsLab/OSL-loc-FigureSkating", revision="main",
    split="fs_comp_test", output_dir="data/figureskating",
    download_format="json",
)

JSON manifests reference shared full broadcasts. Metadata-only reads use annotations_only=True; selective input downloads are also supported. This dataset uses shared raw media and OSL JSON rather than duplicating source videos into separate per-protocol WebDataset archives. Parquet files index the samples for the two Hugging Face configurations; sample_payload contains each OSL row.

Sources and citation

See reports/build.json for the immutable reference commit and source hashes. Research access is manually gated. Source broadcast rights and annotation terms remain with their respective owners; the source code license does not relicense the videos. Cite Hong et al., Video Pose Distillation (ICCV 2021), Hong et al., Spotting Temporally Precise, Fine-Grained Events in Video (ECCV 2022), and Xarles et al., T-DEED (CVPR Workshops 2024).