OpenSportsLab/OSL-cls-tennis-public

OSL Tennis Stroke Classification

17,511 short broadcast-video clips, each labeled with one of four tennis stroke classes. This is a clip-level classification dataset derived from the swing-data localization dataset, with new train/validation splits that hold out whole source matches.

The release contains RGB videos and classification annotations. Skeletons, player tracks, ball/racket detections and model weights are not included. It supports ordinary video classifiers and OpenSportsLib's model-owned frozen pose/tracking frontend with a trainable graph classifier.

Splits and labels

OSL split	Clips	Source matches
train	6,719	15
valid	934	4
test	9,858	9

The classification head is named action; labels are single-label strings. The following class order is used by the supplied OSL configurations:

Class index	Label	Train	Validation	Test
0	far_court_serve	762	95	799
1	far_court_swing	2,611	366	4,145
2	near_court_serve	798	113	779
3	near_court_swing	2,548	360	4,135

near_court and far_court describe the player's side in the broadcast view, not a persistent player identity. swing retains the source's coarse category; this release does not add forehand/backhand or stroke-quality labels. Bounce annotations are excluded. There is no background/no-action class.

Clip construction

• Select only the four serve/swing labels above from localization annotations.
• Request 0.5 seconds before and 0.5 seconds after each selected event.
• Trim bounds at the midpoint to neighboring selected stroke events and at the source video's boundaries. Each clip contains one selected target stroke; it may still contain a bounce or other unselected action.
• Cut at decoded frame boundaries and reset the new clip's timestamps to zero. Source presentation-time spacing is preserved; no temporal interpolation, artificial frame padding or FPS conversion is applied.
• Preserve the decoded source dimensions: 256×144 pixels. Original annotation headers sometimes report 1920×1080; those are not the media's actual dimensions. The converter uses decoded dimensions.
• Encode H.264 MP4 at CRF 18, without audio. This is a lossy video conversion.

Native source rates are 25, 2997/100 frames/s. Measured encoded durations range from 0.434 to 1.001 seconds. Requested windows are at most one second; frame boundaries and source clock precision can make encoded duration slightly longer than one second. Exact source frame/PTS bounds and the target's offset inside the new clip are retained as provenance.

Match-level split policy

The published localization train and validation sets share source matches. This release pools those two sets, samples a fraction of whole matches for validation (validation_fraction=0.2, Python seed 42), and assigns all clips from each held-out match to validation. Test source matches retain their published assignment, subject to the invalid-event exclusions below. No source match appears in more than one split.

Held-out validation matches:

• usopen_2017_mens_round2_federer_youzhny
• usopen_2020_mens_final_zverev_thiem
• usopen_2021_mens_round1_murray_tsitsipas
• usopen_2021_mens_semifinal_djokovic_zverev

The full match assignments and class counts are in reports/conversion.json. These train/validation splits differ from the published localization benchmark; classification and spotting scores should not be compared as if they used identical partitions.

Known invalid source events

3 selected events are excluded because their annotated times lie outside the decoded source video's duration. No source was repaired or silently padded. Details and source hashes are preserved in reports/excluded_events.json.

Source record	Label	Event time (s)	Decoded duration (s)
Tennis_wimbledon_2019_mens_semifinal_federer_nadal_114169_114468	far_court_serve	4.52	4.4
Tennis_wimbledon_2019_mens_semifinal_federer_nadal_114169_114468	near_court_swing	5.36	4.4
Tennis_wimbledon_2019_mens_semifinal_federer_nadal_114169_114468	far_court_swing	7.2	4.4

Storage format

The dataset uses the existing OpenSportsLib Parquet + WebDataset converter and uploader. MP4 bytes are packed into TAR shards rather than published as thousands of individual repository files.

README.md
checksums.sha256
reports/
  conversion.json
  excluded_events.json
train/                       # also valid/ and test/
  metadata.parquet
  shard_manifest.parquet
  shards/shard-000000.tar     # additional shards when required by shard size

metadata.parquet has one row per clip. sample_id is the stable ID, sample_payload is the complete OSL sample as a JSON string, header holds the OSL dataset header, and shard_name/sample_index locate its TAR contents. shard_manifest.parquet maps original relative media paths to wds_member names. For a nine-digit sample key, TAR members are <key>.json (annotations) and <key>.0.mp4 (the single video input).

Each reconstructed OSL sample has:

{
  "id": "<stable-clip-id>",
  "inputs": [{"type": "video", "path": "clips/train/<stable-clip-id>.mp4", "fps": 29.97}],
  "labels": {"action": {"label": "far_court_serve"}},
  "metadata": {
    "base_video": "<source-match>",
    "source_split": "train",
    "source_record_id": "<source-segment>",
    "source_event_index": 0,
    "source_event": {"head": "tennis_action", "label": "far_court_serve", "position_ms": 4520},
    "conversion": {"event_offset_ms": 480, "source_start_frame": 101}
  }
}

This is an abbreviated schema example, not an actual sample. The full conversion block also includes source video/manifest hashes, time base, original frame/PTS bounds, requested and actual intervals, dimensions, frame count and duration.

The Hub's default Parquet configuration exposes metadata, with splits named train, validation, test. Hub validation corresponds to OSL folder valid. It does not automatically decode the MP4s inside the TAR shards:

import json
from datasets import load_dataset

rows = load_dataset("OpenSportsLab/OSL-cls-tennis-public", revision="main", split="train")
sample = json.loads(rows[0]["sample_payload"])
print(sample["labels"]["action"]["label"])

Download and train with OpenSportsLib

Run from an installed OpenSportsLib checkout. The OSL downloader reconstructs ordinary classification JSON and extracts the referenced MP4s:

from pathlib import Path
from opensportslib.tools import download_dataset_split_from_hf
from opensportslib.apis import ClassificationModel, Config

paths = {}
for split in ("train", "valid", "test"):
    result = download_dataset_split_from_hf(
        repo_id="OpenSportsLab/OSL-cls-tennis-public",
        revision="main",
        split=split,
        output_dir="data/tennis-public",
        download_format="parquet",
    )
    paths[split] = Path(result["json_path"]).resolve()

config = Config.from_file("examples/tennis/classification_skeleton.yaml")
overrides = {}
for split, path in paths.items():
    overrides[f"DATA.common.splits.{split}.annotation_path"] = str(path)
    overrides[f"DATA.common.splits.{split}.source_path"] = str(path.parent)
config.update(training={"epochs": 20, "batch_size": 2, "num_workers": 0},
              overrides=overrides)
model = ClassificationModel(config=config)
best_checkpoint = model.train(use_wandb=False)

The skeleton route requires the optional PyG and GPU pose profiles (opensportslib setup --pyg and opensportslib setup --pose). It extracts skeletons online per sampled clip. Detection and pose weights remain frozen; the graph encoder, temporal neck and classifier are trainable. Clip/view tracking state resets for each independent training sample. Skeleton caching is optional and separate from this dataset's media shards.

Evaluation and limitations

Use the held-out validation set for model selection and the test set for final evaluation. Report accuracy, balanced accuracy, macro-F1 and per-class results; swings substantially outnumber serves. This dataset does not establish spotting performance on continuous broadcasts because it contains selected positive stroke clips and supplies no background class.

Source resolution is low, especially for far-court joints. Broadcast cuts, occlusions, camera views and original annotation uncertainty remain present. No new manual relabeling, pose-quality audit or trained-model accuracy benchmark is claimed by this release. Ball/racket motion may carry information that a skeleton-only classifier cannot recover.

Provenance, validation and license declaration

• Localization source: OpenSportsLab/OSL-loc-tennis-public, branch swing-data, immutable commit f1cc2de012aa317883d3dc371953d268f2a2c86b.
• That dataset identifies OpenSportsLab/soccernetpro-localization-tennis as its original videos/annotations source.
• Classification conversion identity: 153c3ef0f834ed6b4c4dc38667bb855c8eceed604be68ace8a24860227abff05.
• Reproduction: tools/convert/tennis_to_classification.py in OpenSportsLib, with default midpoint trimming, seed 42, and --invalid-events skip. Conversion settings, source annotation hashes and release manifest hashes are recorded in reports/conversion.json.
• Every encoded output was decoded to verify frame count, timestamps, dimensions and target inclusion. Publishing verifies byte hashes against the converter's receipts, clip metadata, label counts and source-match separation. checksums.sha256 records original relative MP4 paths and their hashes after extraction; use the entries belonging to each downloaded split.

The localization dataset's source card declares AGPL-3.0; this card carries that declaration forward. Original broadcast footage remains subject to its underlying rights and source terms. This format conversion does not grant additional rights to the footage. Credit the original source dataset and its authors when using these derived clips.