OpenSportsLab/OSL-loc-tennis-public

OSL Tennis Swing Localization

OSL Tennis Swing Localization is a video dataset for tennis action/event localization. It contains 3,445 short broadcast-tennis clips and 35,028 clip-relative event annotations. Labels are the binary swing and sampled no_swing events used by the corresponding OpenSportsLab localization model.

Annotations and sample metadata are published as Parquet for filtering and Dataset Viewer support. Video bytes and matching JSON sidecars are published in split-scoped WebDataset TAR shards.

The corresponding model repository is OpenSportsLab/OSL-loc-tennis-public. This dataset is a separate Hugging Face resource of type dataset; publishing it did not change the model repository.

Dataset structure

Split	Samples/videos	Event annotations	Parquet	WebDataset shards
train	1,368	11,498	1	2
validation	455	3,808	1	1
test	1,622	19,722	1	3
Total	3,445	35,028	3	6

.
├── README.md
├── data/
│   ├── train-00000-of-00001.parquet
│   ├── validation-00000-of-00001.parquet
│   └── test-00000-of-00001.parquet
└── webdataset/
    ├── train-00000.tar
    ├── train-00001.tar
    ├── validation-00000.tar
    ├── test-00000.tar
    ├── test-00001.tar
    └── test-00002.tar

Annotation format

Each Parquet row is one clip. position_ms is relative to the clip start. The nested events list preserves event order and contains:

• head: main_action
• label: swing or no_swing
• position_ms: clip-relative event time in milliseconds
• comment: optional source detail such as serve or forehand_topspin; null when absent

The complete test refinement maps near/far serves and swings to swing, excludes bounce events, and samples one no_swing per swing deterministically with seed 42, scoped by clip ID and at least 251 ms from every swing.

The one-clip refine_test.json smoke-test set is not included. refine_test_full.json, covering all 1,622 test videos, is canonical.

Parquet schema

Field	Type	Description
sample_id	string	Stable sample identifier
split	string	train, validation, or test
video, base_video	string	Clip and source-video identifiers
segment_start_frame, segment_end_frame	int64	Source-video bounds
fps_raw, fps	float64	Measured/source and declared frame rates
width, height	int32	Video dimensions
num_frames, duration_ms	int64	Clip length
num_annotations	int32	Number of nested events
media_type	string	video
video_path	string	Portable original split-relative path
webdataset_key	string	Key shared by MP4 and JSON TAR members
webdataset_shard	string	Repository-relative TAR path
events	list(struct)	Nested localization annotations

No machine-specific absolute paths are stored.

WebDataset organization

Shards are uncompressed TAR archives of approximately 256 MiB and never split a sample. Every video and JSON annotation sidecar share the stable sample ID:

Tennis_<clip-id>.mp4
Tennis_<clip-id>.json

The sidecar contains portable metadata and the complete event list. Parquet columns webdataset_key and webdataset_shard locate the matching pair.

Loading Parquet

from datasets import load_dataset

dataset = load_dataset("OpenSportsLab/OSL-loc-tennis-public")
sample = dataset["train"][0]
print(sample["sample_id"], sample["video_path"])
print(sample["events"])

Direct PyArrow loading:

import pyarrow.parquet as pq

table = pq.read_table("data/train-00000-of-00001.parquet")
print(table.schema)

Streaming WebDataset

import webdataset as wds
from huggingface_hub import hf_hub_url

repo_id = "OpenSportsLab/OSL-loc-tennis-public"
urls = [
    hf_hub_url(repo_id, f"webdataset/train-{index:05d}.tar", repo_type="dataset")
    for index in range(2)
]
sample = next(iter(wds.WebDataset(urls, shardshuffle=False).decode()))
print(sample["__key__"])
video_bytes = sample["mp4"]
annotation = sample["json"]

Validation

Before upload, local validation confirmed complete one-to-one video coverage, unique IDs/paths/events, in-bounds timestamps, matching declared event counts, and no byte-identical video duplicates. It reopened all three Parquet files and all six TAR archives, checking every MP4/JSON pair and matching source, Parquet, and WebDataset counts.

Source and citation

The clips and raw annotations derive from the tennis data released with Spotting Temporally Precise, Fine-Grained Events in Video (ECCV 2022), by James Hong, Haotian Zhang, Michaël Gharbi, Matthew Fisher, and Kayvon Fatahalian.

The upstream six-class test annotations came from OpenSportsLab/soccernetpro-localization-tennis.