MedVidBench: A Benchmark for Medical Video Understanding

Introduced in the paper: MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding

📄 Paper: https://arxiv.org/abs/2512.06581 🌐 Project Page: https://yuhaosu.github.io/MedGRPO/

Version: 11_04 (November 2024) | Test Benchmark

Dataset Description

MedVidBench is a comprehensive test benchmark for evaluating Video Large Language Models (VLMs) on medical and surgical video understanding tasks. This dataset provides questions across 8 diverse tasks (with GPT-4 and Gemini variants for captioning tasks) for public evaluation.

⚠️ IMPORTANT: This public release contains questions only. Ground truth answers are kept private on the evaluation server to ensure fair benchmark comparison.

Key Features

• 6,245 test samples (questions only)
• 8 diverse tasks spanning video, segment, and frame-level understanding
• 8 medical video datasets including AVOS, CholecT50, CholecTrack20, Cholec80_CVS, CoPESD, EgoSurgery, NurViD, and JIGSAWS
• Multi-modal data with video frames and questions
• Structured metadata including FPS, video IDs, temporal spans
• Bounding box annotations for region caption tasks (306 samples)

Tasks

Task	Train	Test	Description
cvs_assessment	-	648	CVS Assessment - Clinical variable scoring
dense_captioning_gemini	-	728	Dense Captioning (Gemini) - Generate detailed descriptions
dense_captioning_gpt	-	751	Dense Captioning (GPT) - Generate detailed descriptions
next_action	-	670	Next Action Prediction - Predict the next surgical step
region_caption_gemini	-	168	Region Caption (Gemini) - Describe specific regions
region_caption_gpt	-	138	Region Caption (GPT) - Describe specific regions
skill_assessment	-	160	Skill Assessment - Evaluate surgical skill levels
stg	-	780	Spatiotemporal Grounding - Locate actions in space and time
tal	-	1,637	Temporal Action Localization - Identify start/end times of actions
video_summary_gemini	-	302	Video Summary (Gemini) - Summarize surgical videos
video_summary_gpt	-	263	Video Summary (GPT) - Summarize surgical videos
Total	-	6,245	8 tasks (11 variants)

Source Datasets

Dataset	Test Samples	Domain
NurViD	2,478	Nursing care videos
CoPESD	1,267	Colonoscopy polyp segmentation
CholecTrack20	795	Cholecystectomy tracking
Cholec80_CVS	648	Cholecystectomy with CVS scores
CholecT50	409	Cholecystectomy triplets
AVOS	321	Abdominal video object segmentation
EgoSurgery	167	Egocentric surgical videos
JIGSAWS	160	Robotic surgery (JHU-ISI)
Total	6,245	8 datasets

Dataset Structure

Data Fields

Each sample contains:

{
    "id": str,                      # Unique identifier (video_id&&start&&end&&fps)
    "video": List[str],             # List of frame file paths
    "sampled_video_frames": List[int],  # Original frame indices
    "conversations": [              # Questions only (no ground truth)
        {"from": "human", "value": str}
    ],
    "qa_type": str,                 # Task type (tal, stg, etc.)
    "dataset_name": str,            # Source dataset
    "train": bool,                  # Train/test split indicator
    "data_source": str,             # Original data source
    "metadata": {                   # Video metadata
        "video_id": str,
        "fps": str,                  # Frames per second
        "input_video_start_time": float,
        "input_video_end_time": float
    },
    "is_RC": bool,                  # Whether sample has region captions
    "RC_info": dict or None         # Bounding box info for region captions
}

Region Caption Task

For region caption samples (is_RC: true):

"RC_info": {
    "start_frame": str,              # Path to frame with bounding box
    "start_frame_bbox": [x1, y1, x2, y2]  # Bounding box coordinates
}

Usage

Loading the Dataset

Note: Only the test split (questions only) is publicly available. Ground truth is kept on the evaluation server.

from datasets import load_dataset
import json

# Load test split
test_data = load_dataset("UIIAmerica/MedVidBench", split="test")

# Or load JSON directly
with open("cleaned_test_data_11_04.json") as f:
    test_data = json.load(f)

# Filter by task
tal_samples = [s for s in test_data if s['qa_type'] == 'tal']
stg_samples = [s for s in test_data if s['qa_type'] == 'stg']

# Filter by dataset
nurvid_samples = [s for s in test_data if s['dataset_name'] == 'NurViD']

Example: Running Inference

import json

# Load test data
with open("cleaned_test_data_11_04.json") as f:
    test_data = json.load(f)

# Run inference on all samples
predictions = []
for sample in test_data:
    # Get question
    question = next(msg['value'] for msg in sample['conversations']
                   if msg['from'] == 'human')

    # Get video frames
    video_frames = sample['video']

    # Your model's prediction
    prediction = your_model.predict(video_frames, question)

    # Store prediction
    predictions.append({
        'id': sample['id'],
        'qa_type': sample['qa_type'],
        'prediction': prediction
    })

# Save predictions for submission
with open("predictions.json", "w") as f:
    json.dump(predictions, f, indent=2)

Submitting for Evaluation

Submit your predictions to the MedVidBench Leaderboard for evaluation.

Evaluation metrics by task:

• TAL: Precision, Recall at IoU 0.3, 0.5, [email protected]
• STG: Spatiotemporal IoU at multiple thresholds
• DVC/VS/RC: BLEU, METEOR, CIDEr, ROUGE-L, LLM Judge scores
• Next Action: Accuracy, per-class average, weighted F1
• Skill/CVS Assessment: Classification accuracy, precision, recall, F1

Version History

• 11_04 (November 2024): Latest version with refinements
• 10_18 (October 2024): Initial release

Citation

If you use this dataset or benchmark, please cite:

@article{su2024medgrpo,
  title={MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding},
  author={Su, Yuhao and Choudhuri, Anwesa and Gao, Zhongpai and Planche, Benjamin and Nguyen, Van Nguyen and Zheng, Meng and Shen, Yuhan and Innanje, Arun and Chen, Terrence and Elhamifar, Ehsan and Wu, Ziyan},
  journal={arXiv preprint arXiv:2512.06581},
  year={2025}
}

📦 Download Test Images

The test set includes 103,742 image frames (~18-19 GB) organized by source dataset:

Option 1: Download ZIP (Recommended)

wget https://huggingface.co/datasets/UIIAmerica/MedVidBench/resolve/main/testdata.zip
unzip testdata.zip

Option 2: Clone with Git LFS

git lfs install
git clone https://huggingface.co/datasets/UIIAmerica/MedVidBench

The testdata/ directory includes:

• All 103,742 test frames organized by dataset (AVOS, NurViD, CAMMA, JIGSAWS, CoPESD, EgoSurgery)
• Organized by source dataset for easy access

Note: The cleaned_test_data_11_04.json file references frame paths. After downloading testdata.zip, update the paths in your code to point to the extracted frames.

🔗 Links

• 📄 Paper: arXiv:2512.06581
• 🌐 Project Page: yuhaosu.github.io/MedGRPO
• 💻 GitHub: github.com/YuhaoSu/MedGRPO
• 🤗 Dataset: huggingface.co/datasets/UIIAmerica/MedVidBench
• 📊 Leaderboard: huggingface.co/spaces/UIIAmerica/MedVidBench-Leaderboard

License

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made
• NonCommercial — You may not use the material for commercial purposes
• ShareAlike — If you remix, transform, or build upon the material, you must distribute your contributions under the same license

Note: Individual source datasets (AVOS, CholecT50, NurViD, etc.) may have additional restrictions. Please refer to their original licenses and comply with all applicable terms.

Dataset Preparation

This benchmark was prepared using:

1. Video sampling at variable FPS (0.1-1.0 fps) based on video length
2. LLM-based caption generation (GPT-4, Gemini)
3. Human verification and cleaning
4. Structured ground truth annotation (kept private for fair evaluation)

For research inquiries about the full dataset with ground truth, please contact the authors.

Contact

For questions or issues, please:

• Open an issue on the GitHub repository
• Visit the project page
• Email: [Contact information from paper]