Dataset Card for "MM-OR-SG-Dataset"

Description

MM-OR-SG-Dataset is a curated subset of the MultiModal Operating Room Dataset for Semantic Understanding of High-intensity Surgical Environments, made for tiny projects on two tasks: (1) scene graph generation and (2) event significance classification.

Each sample includes three consecutive surgical video frames and a corresponding panoptic segmentation mask, all resized to 256×256. The annotation for each sample provides structured scene graph relations and an event significance label.

Note: The event significance annotation ("routine", "noteworthy", "critical") is not part of the original MM-OR dataset, and is Pseudo-labeled with an SVC.

Dataset Structure

• images: Sequence of four images per sample (three RGB frames + one panoptic segmentation mask, all 256×256).
• annotations: A JSON string containing:
  • rel_annotations: List of (subject, relation, object) scene graph triplets.
  • event_significance: One of "routine", "noteworthy", or "critical" (added by this work).

Usage Example

from datasets import load_dataset

ds = load_dataset("fiqqy/MM-OR-SG-Dataset")
sample = ds["train"][0]
images = sample["images"]        
import json
annotations = json.loads(sample["annotations"])
print("Scene Graph Relations:", annotations["rel_annotations"])
print("Event Significance:", annotations["event_significance"])

If you use this dataset, please cite the original MM-OR paper:

@inproceedings{ozsoy2025mm,
  title={Mm-or: A large multimodal operating room dataset for semantic understanding of high-intensity surgical environments},
  author={{\"O}zsoy, Ege and Pellegrini, Chantal and Czempiel, Tobias and Tristram, Felix and Yuan, Kun and Bani-Harouni, David and Eck, Ulrich and Busam, Benjamin and Keicher, Matthias and Navab, Nassir},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={19378--19389},
  year={2025}
}