EgoExOR-HQ: An Ego-Exo-Centric Operating Room Dataset for Surgical Activity Understanding

EgoExOR-HQ — This repository hosts the enriched high-quality release of the EgoExOR dataset. For scene graph generation code, benchmarks, and pretrained models, see the main EgoExOR repository.

Authors: Ege Özsoy, Arda Mamur, Felix Tristram, Chantal Pellegrini, Magdalena Wysocki, Benjamin Busam, Nassir Navab

✨ What's New in EgoExOR-HQ

This release adds:

• High-quality images — 1344×1344 resolution (instead of 336×336)
• Raw depth images — From external RGB-D cameras (instead of pre-merged point clouds), so you can build merged or per-camera point clouds for your use case
• Per-device audios — Separate audio streams per microphone

Overview

Operating rooms (ORs) demand precise coordination among surgeons, nurses, and equipment in a fast-paced, occlusion-heavy environment, necessitating advanced perception models to enhance safety and efficiency. Existing datasets either provide partial egocentric views or sparse exocentric multi-view context, but do not explore the comprehensive combination of both.

We introduce EgoExOR, the first OR dataset and accompanying benchmark to fuse first-person and third-person perspectives. Spanning 94 minutes (84,553 frames at 15 FPS) of two emulated spine procedures—Ultrasound-Guided Needle Insertion and Minimally Invasive Spine Surgery—EgoExOR integrates:

• Egocentric: RGB, gaze, hand tracking, audio from wearable glasses
• Exocentric: RGB and depth from RGB-D cameras, ultrasound imagery
• Annotations: 36 entities, 22 relations (568,235 triplets) for scene graph generation

This dataset sets a new foundation for OR perception, offering a rich, multimodal resource for next-generation clinical perception.

🌟 Key Features

• Multiple modalities — RGB video, audio (full waveform + per-frame snippets, per-device), eye gaze, hand tracking, raw depth, and scene graph annotations
• Time-synchronized streams — All modalities aligned on a common timeline for precise cross-modal correlation
• High-resolution RGB — 1344×1344 frames for fine-grained visual analysis
• Raw depth — Build custom point clouds or depth-based models; depth from external RGB-D cameras only
• Per-device audio — Separate microphone streams for spatial or multi-channel audio processing

📂 Dataset Structure

The dataset is distributed as phase-level HDF5 files for efficient download:

File	Description
miss_1.h5	MISS procedure, phase 1
miss_2.h5	MISS procedure, phase 2
miss_3.h5	MISS procedure, phase 3
miss_4.h5	MISS procedure, phase 4

To obtain a single merged file (including splits), use the merge utility from the main EgoExOR repository (see data/README.md).

HDF5 Schema

/metadata
  /vocabulary/entity        — Entity names and IDs (instruments, anatomy, etc.)
  /vocabulary/relation      — Relation names and IDs (holding, cutting, etc.)
  /sources/sources          — Camera/source names and IDs (head_surgeon, external_1, etc.)
  /dataset                  — version, creation_date, title

/procedures/{procedure}/phases/{phase}/takes/{take}/
  /sources                  — source_count, source_0, source_1, … (camera roles)
  /frames/rgb               — (num_frames, num_cameras, H, W, 3) uint8 — 1344×1344
  /eye_gaze/coordinates     — (num_frames, num_ego_cameras, 3) float32 — gaze 2D + camera ID
  /eye_gaze_depth/values    — (num_frames, num_ego_cameras) float32
  /hand_tracking/positions  — (num_frames, num_ego_cameras, 17) float32
  /audio/waveform           — Full stereo waveform
  /audio/snippets           — 1-second snippets aligned to frames
  /audio/per_device/        — Per-microphone waveform and snippets
  /point_cloud/depth/values — Raw depth images (external cameras; others zero-filled)
  /point_cloud/merged/      — Not populated; use raw depth to build point clouds yourself
  /annotations/             — Scene graph annotations (frame_idx, rel_annotations, scene_graph)

/splits
  train, validation, test   — Split tables (procedure, phase, take, frame_id)

Note: Camera/source IDs in eye_gaze/coordinates map to metadata/sources for correct source names.

⚙️ Efficiency and Usability

• HDF5 — Hierarchical structure, partial loading, gzip compression
• Chunking — Efficient access to frame ranges for sequence-based training
• Logical layout — procedures → phases → takes → modality for easy navigation

📜 License

Released under the Apache 2.0 License. Free for academic and commercial use with attribution.

📚 Citation

@misc{özsoy2025egoexoregoexocentricoperatingroom,
      title={EgoExOR: An Ego-Exo-Centric Operating Room Dataset for Surgical Activity Understanding}, 
      author={Ege Özsoy and Arda Mamur and Felix Tristram and Chantal Pellegrini and Magdalena Wysocki and Benjamin Busam and Nassir Navab},
      year={2025},
      eprint={2505.24287},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.24287}, 
}

🔗 Related Resources

• Original EgoExOR (v1) — ardamamur/EgoExOR — 336×336 images, pre-merged point clouds, merged audio
• Code, benchmarks, pretrained model — github.com/ardamamur/EgoExOR

---

Dataset: TUM/EgoExOR · Last Updated: February 2025