🏃 AthletePose3D: A Benchmark Dataset for 3D Human Pose Estimation and Kinematic Validation in Athletic Movements (CVSports at CVPR 2025)

📌 Overview

AthletePose3D (AP3D) is a novel dataset for monocular 3D human pose estimation in sports biomechanics, designed to capture high-speed, high-acceleration movements. Alongside the raw dataset, we also provide a training-ready version prepared for 2D and 3D pose estimation modeling, including both preprocessed annotations and AP3D fine-tuned model parameters.

To download, please read the license agreement.

📂 Dataset Features

• 🏅 12 sports motions across various disciplines
• 🎞️ 1.3M frames & 165K postures
• ⚡ Focus on high-intensity athletic movements

📊 Model Evaluation

• 📉 SOTA models trained on conventional datasets struggle with athletic motions
• 🎯 Fine-tuning on AthletePose3D reduces MPJPE from 214mm → 65mm (69% improvement!)

🔬 Kinematic Validation

• ✅ Strong joint angle correlation
• ⚠️ Limitations in velocity estimation

🚀 Contribution

• Benchmarking monocular pose estimation for sports
• Advancing pose estimation in high-performance environments

💡 Example

example.mp4

📂 Dataset Structure

• /AthletePose3D/
  • /data/ (video and motion data)
    • /train_set/
      • /S1/ (subject)
        • Axel_1_cam_1.mp4 (video file)
        • Axel_1_cam_1.json (video and motion information)
        • Axel_1_cam_1.npy (motion data)
        • Axel_1_cam_1_coco.npy (COCO keypoints)
        • Axel_1_cam_1_h36m.npy (H3.6M keypoints)
      • /S2/
      • ...
    • /valid_set/
    • /test_set/
  • /pose_2d/ (2D pose estimation ready data)
    • /annotations/ (Annotations in COCO Format)
      • train_set.json
      • ...
    • /det_result/ (Detected with YOLOv8)
      • ap2d_train_det.json
      • ...
    • /train_set/ (Image files)
    • /valid_set/
    • /test_set/
  • /pose_3d/ (3D pose estimation ready data)
    • /frame_81/
    • train.pkl
    • valid.pkl
  • cam_param.json (camera parameters)

⬇️ Download AthletePose3D

The dataset is available for download at the following link: Download AthletePose3D

📖 Reference

Please consider citing our work if you find it helpful to yours:

@misc{yeung2025athletepose3d,
      title={AthletePose3D: A Benchmark Dataset for 3D Human Pose Estimation and Kinematic Validation in Athletic Movements}, 
      author={Calvin Yeung and Tomohiro Suzuki and Ryota Tanaka and Zhuoer Yin and Keisuke Fujii},
      year={2025},
      eprint={2503.07499},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07499}, 
}

📄 License

For non-commercial and scientific research purposes only. For details refer to LICENSE