Kingston QTM C3D — Ground Truth & v11e Pose Models

Complete dataset for validating markerless pose estimation against Qualisys marker-based motion capture. Collected at Kingston University London with 9 athletes performing 12 sport science drills.

Downloads

Archive	Contents	Compressed	Extracted
`kingston-v11e-complete.tar.gz`	Everything (ground truth + results + models)	2.0 GB	5.3 GB
`kingston-ground-truth.tar.gz`	Qualisys C3D + QTM files only	1.0 GB	2.7 GB
`kingston-v11e-results.tar.gz`	v11e inference results only (FP32 + FP16)	223 MB	1.7 GB

Quick Start

Download everything

huggingface-cli download torouni/kingston-qtm-c3d-models kingston-v11e-complete.tar.gz --local-dir .
tar -xzf kingston-v11e-complete.tar.gz

Download just the results

huggingface-cli download torouni/kingston-qtm-c3d-models kingston-v11e-results.tar.gz --local-dir .
tar -xzf kingston-v11e-results.tar.gz

Download ground truth only

huggingface-cli download torouni/kingston-qtm-c3d-models kingston-ground-truth.tar.gz --local-dir .
tar -xzf kingston-ground-truth.tar.gz

Download individual files (Python)

from huggingface_hub import hf_hub_download

# Single model file
hf_hub_download("torouni/kingston-qtm-c3d-models", "student_v11e_52kp_fp16.onnx.data", local_dir=".")

# Single result file
hf_hub_download(
    "torouni/kingston-qtm-c3d-models",
    "results/drills-student-v11e-fp16/cmj/p01-cmj-trial01_results_rtdetr-student-v11e.json",
    local_dir="."
)

Browse and download by category

# All inference results (browsable individual files)
huggingface-cli download torouni/kingston-qtm-c3d-models --include "results/*" --local-dir .

# FP16 results only
huggingface-cli download torouni/kingston-qtm-c3d-models --include "results/drills-student-v11e-fp16/*" --local-dir .

Repository Contents

torouni/kingston-qtm-c3d-models/
│
│  Archives
├── kingston-v11e-complete.tar.gz       # Full package (2.0 GB)
├── kingston-ground-truth.tar.gz        # C3D + QTM ground truth (1.0 GB)
├── kingston-v11e-results.tar.gz        # Inference results only (223 MB)
│
│  Model weights
├── student_v11e_52kp_fp16.onnx.data    # v11e FP16 weights (174 MB)
│
│  Browsable inference results
├── results/
│   ├── drills-student-v11e/            # 191 FP32 JSONs (1.1 GB)
│   └── drills-student-v11e-fp16/       # 191 FP16 JSONs (632 MB)
│
└── README.md

Archive structures

kingston-v11e-complete.tar.gz           kingston-v11e-results.tar.gz
├── ground-truth/                       ├── v11e-fp32-results/
│   ├── c3d/P01-P09/  (198 files)       │   ├── 10-5-5-ball/
│   └── qtm/P01-P09/  (198 files)       │   ├── 10m-sprint/
├── results/                            │   ├── 5-0-5-left/
│   ├── drills-student-v11e/            │   ├── 5-0-5-right/
│   └── drills-student-v11e-fp16/       │   ├── 5-10-5/
├── models/                             │   ├── broad-jump/
│   ├── student_v11e_52kp_fp16.onnx     │   ├── cmj/
│   ├── student_v11e_52kp_fp16.onnx.d…  │   ├── diamond-dribble-left/
│   ├── student_v11a_52kp_fp16.onnx     │   └── diamond-dribble-right/
│   └── student_v11a_52kp.onnx          └── v11e-fp16-results/
└── README.md                               └── (same structure)

kingston-ground-truth.tar.gz
├── c3d/P01-P09/  (198 C3D files)
└── qtm/P01-P09/  (198 QTM files)

Dataset Details

Property	Value
Institution	Kingston University London
Participants	9 athletes (P01–P09)
Drills	12 sport science drills
Total sessions	198
Ground truth system	Qualisys marker-based motion capture
Ground truth formats	C3D, QTM
Total size (uncompressed)	5.3 GB

Drills

Drill	Category	Description
10-5-5 Ball	Agility	Shuttle run with ball (10m-5m-5m)
10m Sprint	Speed	Maximal 10-meter sprint
5-0-5 Left	Agility	Change of direction (left)
5-0-5 Right	Agility	Change of direction (right)
5-10-5	Agility	Pro agility shuttle
Broad Jump	Power	Standing broad jump
CMJ	Power	Counter-movement jump
Diamond Dribble Left	Agility	Diamond pattern dribble (left)
Diamond Dribble Right	Agility	Diamond pattern dribble (right)

Model Performance

Evaluated against Qualisys ground truth on the Kingston dataset:

Model	PA-MPJPE (mm)	vs 3DAT
v11e FP16	39.6	32% better
v11a FP16	41.2	29% better
3DAT (production)	58.2	baseline

Model Architecture

Base: ViTPose (Vision Transformer for Pose Estimation)
Training: Knowledge distillation from SynthPose Teacher (ViTPose-L)
Keypoints: 52 (17 COCO + 35 biomechanical markers)
Input: 256x192 RGB crop → [1, 3, 256, 192]
Output: 52 heatmaps → [1, 52, 128, 96]
Format: ONNX (opset 17), FP16 internal computation with FP32 I/O

Inference Result Format

Each JSON contains per-frame, per-athlete pose predictions:

{
  "video_info": { "input": "path/to/video.mp4" },
  "model_info": {
    "model_name": "rtdetr-student-v11e-fp16",
    "num_keypoints": 52,
    "num_coco": 17,
    "heatmap_size": [128, 96],
    "input_size": [256, 192]
  },
  "frames": [
    {
      "frame_idx": 0,
      "athletes": [
        {
          "bbox": [x1, y1, x2, y2],
          "keypoints": {
            "nose": {"x": 320.5, "y": 180.2, "confidence": 0.92},
            "left_shoulder": {"x": 305.1, "y": 220.8, "confidence": 0.95}
          },
          "keypoints_52": {
            "sternum": {"x": 312.0, "y": 260.5, "confidence": 0.88},
            "C7": {"x": 311.2, "y": 195.3, "confidence": 0.85}
          }
        }
      ]
    }
  ]
}

52 Keypoint Layout

COCO (1–17): nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles

Biomechanical (18–52): sternum, shoulder markers, lateral/medial elbow, lateral/medial wrist, ASIS/PSIS, lateral/medial knee, lateral/medial ankle, 5th metatarsal, toe, big toe, calcaneus, C7, L2, T11, T6

Citation

@misc{kingston_qtm_c3d_2026,
  title={Kingston Sport Science Ground Truth Dataset for Markerless Pose Estimation Validation},
  author={Nelson AI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/torouni/kingston-qtm-c3d-models}
}

License

CC BY-NC 4.0 — free for research and non-commercial use.

torouni
/

kingston-qtm-c3d-models