nosuke113/caltennis-human-stroke-joints

CalTennis Human 3D Stroke Joints

Related releases: companion G1 references https://huggingface.co/datasets/nosuke113/caltennis-g1-hitting-references · code https://github.com/humanoid-tennis-project/g1-tennis-swing

3D trajectories of 22 body joints for tennis strokes, estimated with GVHMR from the public CalTennis videos. This is the human-side source of the companion CalTennis-G1 Hitting References release (same anonymized clip_id).

• 544 clips (forehand 245, backhand 140, serve 47, volley 84, overhead 28), 30 Hz, 0.97-9.50 s (median 2.47 s), 43,474 frames = 24.15 min, about 13 MB.
• 22 recordings from 10 CalTennis sessions.
• 519 clips have one hand-annotated hit (racket-ball contact) time.

This is a derived, unofficial dataset, not endorsed by the CalTennis, GVHMR or SMPL-X authors.

Joint positions only. SMPL-X shape coefficients (betas), SMPL-X pose parameters, camera parameters, meshes, model files and video frames are deliberately not included (see License).

Intended use

Motion analysis, retargeting to characters or robots, and reference motions for motion-tracking research, for non-commercial purposes. Not a measurement of the players; not suitable for biomechanical, clinical or forensic conclusions. Do not use it to identify or profile the people in the source videos.

Files

metadata.jsonl          one row per clip (544)
data/clip_XXXX.npz
recordings.json         CalTennis source video for each recording_id (path, sha256, pinned URL)

key	shape	meaning
fps	()	30
joint_names	(22,)	SMPL-X body joint order (pelvis, hips, spine, knees, ankles, feet, neck, collars, head, shoulders, elbows, wrists)
joints_global	(T, 22, 3)	joint positions, m, gravity-aligned world frame, Y-up (GVHMR "global" output)
source_time_s	(T,)	time from the start of the source video, s
source_frame_30hz	(T,)	frame index in the source video on a 30 Hz grid
person_detected	(T,)	false where the person box was interpolated rather than detected

metadata.jsonl: clip_id, file, sha256, stroke_type, frames, fps, duration_s, recording_id, source_frame_start_30hz, source_frame_end_30hz, hit_annotated, hit_events (local_frame, clip_time_s, source_frame_30hz, evidence = visible_contact | estimated_swing, uncertainty_frames_30hz), in_g1_stage1_set, and simple diagnostics (direct_detection_fraction, interpolated_tracking_frames, max_root_step_m).

import json, numpy as np
rows = [json.loads(l) for l in open("metadata.jsonl", encoding="utf-8")]
r = next(r for r in rows if r["stroke_type"] == "backhand" and r["hit_annotated"])
with np.load(r["file"], allow_pickle=False) as m:
    j = m["joints_global"]   # (T, 22, 3)
print(j.shape, r["hit_events"][0]["local_frame"])

How it was built

1. CalTennis videos (revision 4af17dd4d398aad4116a35b5a74de4388658d356) were cut into stroke intervals.
2. GVHMR (rev ee960bb6) estimated SMPL-X neutral body motion for the tracked player; the 22 body joints were computed by forward kinematics and only those positions are released. Camera intrinsics were GVHMR defaults (image-center principal point), not CalTennis calibration.
3. Clips were visually selected (544) and labeled with stroke type; one annotator marked the hit time on the source video (519 clips; serve included).

Known issues and limitations

• Monocular estimates from far cameras (about 13-17 m): depth, foot contact and fast wrist motion have errors; feet slide on the ground (no contact correction is applied here).
• Joint positions implicitly contain the estimated limb lengths of the person.
• 15 clips contain interpolated person boxes; 4 have direct detection below 95%. A few clips show large root-rotation jumps between adjacent frames.
• Hit times: single annotator, no agreement study; estimated_swing events are estimates. No ball or racket data.
• No player labels and no official split; recordings are unevenly represented.

License and attribution

CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/), following the source videos. Changes: 3D pose estimation, joint extraction, clipping, stroke labels and hit annotations added.

• Source videos: CalTennis by Ilona Demler, Xinran Xie, Blake Werner, Anna Szczuka, Pietro Perona (Caltech), CC BY-NC 4.0, https://huggingface.co/datasets/demalenk/caltennis, https://arxiv.org/abs/2606.20542.
• Pose estimation: GVHMR, Zhejiang University, https://github.com/zju3dv/GVHMR.
• Body model: joint positions were computed with SMPL-X (Max Planck Institute for Intelligent Systems, https://smpl-x.is.tue.mpg.de). No SMPL-X model or shape space is distributed. Please also cite SMPL-X (Pavlakos et al., "Expressive Body Capture: 3D Hands, Face, and Body from a Single Image", CVPR 2019) and GVHMR (Shen et al., "World-Grounded Human Motion Recovery via Gravity-View Coordinates", SIGGRAPH Asia 2024).

@article{caltennis2026,
  author  = {Demler, Ilona and Xie, Xinran and Werner, Blake and Szczuka, Anna and Perona, Pietro},
  title   = {CalTennis: Large Multi-View Tennis Video Dataset and Benchmark of Monocular-to-3D Pose Estimation},
  journal = {arXiv:2606.20542},
  year    = {2026}
}

---

日本語の概要

CalTennisの公開映像からGVHMRで推定した、テニスの打球動作544クリップ（30Hz、計24.15分）の身体22部位の3D位置です。打球種別と、519クリップには人手で付けた打撃時刻があります。G1版データセットと同じclip_idで対応します。

公開するのは関節位置のみです。SMPL-Xの形状係数（betas）・姿勢パラメータ・カメラ情報・メッシュ・モデルファイル・映像は含みません。映像推定のため奥行きや足の接地に誤差があり、足滑りの補正はしていません。ライセンスはCC BY-NC 4.0（非営利）です。CalTennis・GVHMR・SMPL-Xを引用してください。