Complete 3,000-clip, four-event-class spotting release prepared for OpenSportsLib and T-DEED reproduction. The original FineDiving AQA taxonomy has 52 action classes; this spotting protocol uses Entry, Som(s).Pike, Som(s).Tuck and Twist(s).

Split	Clips
Train	1,801
Validation	450
Test	749

Media and timing

Videos are FFV1 RGB-lossless Matroska files, packaged as OSL WebDataset TAR shards with Parquet metadata. Every output frame was fully decoded and its RGB pixels verified against the decoded source JPEG sequence. There are 312,256 frames in total. Original JPEGs remain available in sources/frames.zip.

25 fps is an explicit conversion clock, not a measured camera frame rate. Upstream labels use fps: -1. Original zero-based event frame indices are retained; position_ms = frame * 40. Each record retains the upstream record, source JPEG paths, lossless-video checksum, and decoded-RGB checksum. Use frame-based tolerances for the precise event spotting benchmark.

This replaces an older metadata-only conversion whose video inputs were absent, FPS was zero, and event timestamps were all zero. That conversion remains in repository history; it must not be used for reproduction. The raw branch retains the earlier native source annotations.

Loading

from opensportslib.tools import download_dataset_split_from_hf
download_dataset_split_from_hf(
    repo_id="OpenSportsLab/OSL-loc-FineDiving", revision="main",
    split="test", output_dir="data/finediving", download_format="parquet",
)

Authentication and accepted repository access conditions are required. The downloaded split contains a reconstructed OSL JSON manifest and unpacked videos. Dataset Viewer rows contain OSL sample_payload JSON strings; they are not direct Hugging Face Video columns. Original frame-indexed annotations are under annotations/. See reports/build.json for validation and source hashes.

Sources and terms

Original dataset: https://github.com/xujinglin/FineDiving
Spotting labels: https://github.com/arturxe2/T-DEED/tree/d953b4de8a2ad941c565e1683c5b07255cbe644e/data/finediving
Source frame mirror: https://huggingface.co/datasets/ZWJ666/FineDiving/tree/56adbfbc482ce5eda2aaa1033d7a3212613b6cfc
The mirror traces the frame archive to OpenDataLab/FineDiving. The archive's SHA-256 is recorded in the build report.

FineDiving's research and attribution terms apply; the software repository's MIT license does not replace the dataset release agreement. OpenSportsLab confirmed permission to mirror the dataset and annotations for this publication. The existing access gate is preserved. Cite Xu et al., FineDiving (CVPR 2022), Hong et al., Spotting Temporally Precise, Fine-Grained Events in Video (ECCV 2022), and Xarles et al., T-DEED (CVPR Workshops 2024).

Downloads last month: 9

Total file size:

30.3 GB