Trauma THOMPSON Dataset

How to Access the Dataset

To ensure responsible use, access to the dataset is granted after agreeing to our terms of use.

Please complete the user consent form at the following link.

Upon submission of the form, you will be automatically redirected to a page containing the download links for the dataset.

📦 Dataset Description

The Trauma THOMPSON Dataset is designed for multimodal research in egocentric video analysis of life-saving medical procedures. It supports multiple tasks including action recognition, hand tracking, object detection, and visual question answering.

---

🔍 Action Recognition & Anticipation

• 177 videos of standard (regular) emergency procedures
• 43 videos of just-in-time (JIT) procedures

---

✋ Hand Tracking

• 30 videos (subset of the regular procedure dataset)
• Annotations include:
  • Hand bounding boxes in COCO format
  • Left/right hand identification

---

🛠 Object Detection

• 25,000 frames
• 12 common surgical tools
• Bounding box annotations for tool presence and location in YOLO format

---

❓ Visual Question Answering (VQA)

• 600,000 frames with corresponding VQA annotations

---

📁 Dataset Structure

1. Action Recognition & Anticipation

Input: Pre-extracted video frames
Folder Structure:

actions/  
├── videos/  
│   ├── P01_01_00/  
│   │   ├── img_00001.jpg  
│   │   ├── img_00002.jpg  
│   ├── P01_01_01/  
│   │   ├── img_00001.jpg  
│   │   ├── img_00002.jpg  
│   ├── ...  
├── annotations.csv  

Annotations:

• annotations.csv contains video-level labels (e.g., procedure type, timestamps).

---

2. Hand Tracking

Input: Full videos + per-frame bounding boxes
Folder Structure:

hands/  
├── videos/  
│   ├── P01_01.mp4  
│   ├── P01_02.mp4  
│   ├── ...  
├── bbx/  
│   ├── P01_01/  
│   │   ├── P01_01_00001.json  
│   │   ├── P01_01_00002.json  
│   ├── P01_02/  
│   │   ├── P01_02_00001.json  
│   │   ├── P01_02_00002.json  
│   ├── ...  

Annotation Format (COCO-style JSON):

{
    "categories": [
        {"id": 0, "name": "left hand", "supercategory": "hand"},
        {"id": 1, "name": "right hand", "supercategory": "hand"}
    ],
    "images": [
        {
            "id": "frame_number",
            "file_name": "P01_01_00001.jpg",
            "height": "h",
            "width": "w",
            "channel": 3
        }
    ],
    "annotations": [
        {
            "id": "annotation_id",
            "image_id": "frame_number",
            "category_id": 0, 
            "bbox": ["x", "y", "w", "h"],
            "area": "w * h",
            "iscrowd": 0
        }
    ]
}

---

3. Object Detection (Surgical Tools)

Input: Single frames + tool annotations
Folder Structure:

objects/  
├── images/  
│   ├── P01_01_00001.jpg  
│   ├── P01_01_00002.jpg  
│   ├── ...  
├── labels/  
│   ├── P01_01_00001.txt  
│   ├── P01_01_00002.txt 
│   ├── ...  

Annotation Format (YOLO-style):

• Each .txt file contains:

  <object-class> <x-center> <y-center> <width> <height>
  

---

4. Visual Question Answering (VQA)

Input: Frames + question-answer pairs
Folder Structure:

vqa/  
├── images/  
│   ├── P01_01_00001.jpg  
│   ├── P01_01_00002.jpg  
│   ├── ...  
├── questions.json  
├── annotations.json  

Question Format (questions.json):

{
    "info": {
        "description": "Trauma THOMPSON VQA dataset",
        "version": "1.0",
        "year": 2025
    },
    "data_type" : "mscoco",
    "license" : {
        "url": "https://creativecommons.org/licenses/by-nc-sa/4.0/",
        "name": "CC BY-NC-SA 4.0"
    },
    "questions": [
        {
            "image_name": "P01_01_00001.jpg",
            "question": "Is the patient bleeding?",
            "question_id": 10100001001
        }
    ]
}

Answer Format (annotations.json):

{
    "info": {
        "description": "Trauma THOMPSON VQA dataset",
        "version": "1.0",
        "year": 2025
    },
    "data_type" : "mscoco",
    "license" : {
        "url": "https://creativecommons.org/licenses/by-nc-sa/4.0/",
        "name": "CC BY-NC-SA 4.0"
    },
    "annotations": [
        {
            "question_id": 10100001001,
            "image_name": "P01_01_00001.jpg",
            "answers": [
                {"answer": "Yes", "confidence": "high", "answer_id": 1}
            ]
        }
    ]
}

---

📌 Key Features

✅ Multi-task annotations (action, hands, tools, VQA)
✅ Structured folder hierarchy for easy data loading
✅ Standard formats (COCO for hands, YOLO for tools, VQA-JSON)

---

📜 License

CC BY 4.0