Bifocal-OR contains identifiable video and audio of the people who took part in the recordings. They consented to the publication of the recordings for scientific research. To access the data, please agree to the terms below.

Bifocal-OR: A Dual-Scale Multi-Sensor Dataset for Ophthalmic Surgery

Bifocal-OR links what the surgical team does in the operating room with what happens under the microscope. It contains nine realistically simulated ophthalmic procedures (epiretinal membrane peeling on eye phantoms and an ex-vivo porcine eye), recorded with five room-wide RGB-D cameras, a microscope-mounted camera, an ambient room microphone and wearable microphones for each team member. Unscripted events such as sensor dropouts, delays, visitors and role changes were kept.

Paper: TODO arXiv link · Code: TODO GitHub link · Project page: https://ash-tum.github.io/Bifocal-OR/

DRAFT – sections marked TODO are completed once the timing of all streams is verified.

At a glance


Runs	9 (Runs 1–8 plastic eye phantoms, Run 9 ex-vivo porcine eye), 31–45 min each, 5.6 h in total
Room cameras	5 × Orbbec Femto Mega, synchronized frames: color JPEG 1920 × 1080, depth 16-bit TIFF in millimetres 640 × 576 (NFOV); 7.5 frames/s nominal (7.48 on average), with a timestamp for every frame; 14,000–21,000 frames per camera and run
Microscope	Basler camera in the microscope's optical path, 4024 × 3036, H.265 MP4 stored at 29 fps but captured at about 21–29 frames/s (real frame times provided), with the microphone recorded alongside
Audio	1 ambient room microphone (recorded by the microscope PC) + up to 6 wearable smartphone microphones, synchronized, 48 kHz / 16-bit WAV
Size	about 520 GB (color frames 351 GB, depth frames 120 GB, microscope 29 GB, audio 12 GB)
Phase labels	5 phases, all runs (19,084 labeled seconds)
Speaker labels	manually audited: Runs 3, 4, 7 · automatic pseudo-labels: Runs 1–9
Calibration	factory intrinsics + 3 extrinsic calibrations (1.6 / 2.7 / 4.3 px RMS)
Fiducials	ArUco markers on every team member (head, front, back) and on 4 mobile objects
License	CC BY-NC 4.0, gated access for research use

Repository structure

README.md                    this file
LICENSE.md                   license and terms of use
VALIDATION.md                automatic checks of the annotation files against the paper
annotations/
  phases/runNN.csv           phase segments per run (+ all_runs.csv)
  speakers/audited/runNN.*   manually audited speaker activity, Runs 3, 4, 7 (.rttm and .csv)
  speakers/pseudo/runNN.*    automatic speaker activity, Runs 1-9 (.rttm and .csv, + diagnostics/)
  baselines/pyannote_ambient/runNN.*   far-field diarization baseline of the paper, Runs 3, 4, 7
calibration/
  README.md                  conventions and file formats
  intrinsics/camera0X.json   factory calibration as stored by the recording software
  intrinsics_opencv.json     the same, converted to OpenCV matrices and distortion vectors
  extrinsics/*.json          charuco_dist (default), after, shipped
  charuco_board.json         ChArUco board of the recording setup
  charuco_recording/         the recording of the empty room with the ChArUco board (frames, timestamps)
metadata/
  runs.csv                   one row per run: date, take, tissue, split, durations, notes
  audio_channels.csv         which wearable microphone worked in which run
  events.csv                 milestones and notes from the director's log (director's clock, approximate)
  aruco_markers.csv          marker IDs per role and object
  workflow_playbook.xlsx/csv role-by-role workflow script (Table S1 of the paper)
sync/
  README.md                  how the time bases fit together
  offsets.csv                per-run start of the audio and of the microscope clock on the room-camera timeline
  room_frame_times/runNN.csv time of every room-camera frame (seconds since the first frame)
  microscope_frame_times/    real time of the microscope video frames (from the encoder log)
runs/runNN/
  room/color/camera0X.tar    color frames of one camera (color_<frame>_camera0X.jpg)
  room/depth/camera0X.tar    depth frames of one camera (depth_<frame>_camera0X.tiff, 16-bit, millimetres, 0 = no depth)
  room/timestamps.arrow      capture timestamps of every frame (Apache Arrow)
  room/exported_files.arrow  frame index to file name
  room/calibration/          factory calibration of the five cameras for this run
  microscope/microscope.mp4  microscope video (Run 1: microscope_part1.mp4 and _part2.mp4)
  microscope/microscope_audio.wav   microphone recorded with the microscope video
  audio/<role>.wav           synchronized audio: room, surgeon, ota, runner, anesthesiologist,
                             nurse, spectator (only channels that recorded)

Frame numbers are shared across the five cameras of a run (synchronized export); the capture time of every frame is in room/timestamps.arrow. Unpack a camera with tar -xf camera01.tar.

Time bases and synchronization

Room-camera timeline: t = 0 is the first exported frame of the run. The time of every frame is in sync/room_frame_times/runNN.csv (and room/timestamps.arrow). Frames are 133.4 ms apart, with occasional single missing frames, so use the table rather than a constant frame rate.
Phase labels are in seconds on the room-camera timeline.
Audio: the wearable tracks were aligned to the ambient microphone with the claps at the start and end of every run (one offset plus a linear clock-drift correction per track), then refined by waveform alignment (TODO: final wording). The residual offset between the synced tracks is below about 65 ms at all checked points. The synced audio starts at offset_s (row audio in sync/offsets.csv); speaker labels are in audio time = t − offset_s.
Microscope: the video is stored at 29 frames/s, but frames were captured at a lower, varying rate. sync/microscope_frame_times/ gives the real time of the frames; the microscope clock starts at offset_s (row microscope).

Annotations

Surgical phases

Five macroscopic phases, one contiguous segment each per run (phase_id, phase, start_s, end_s, duration_s). Boundaries follow the workflow playbook:

ID	Phase	Starts with
0	Room_Prep	start of room preparation (see playbook)
1	Patient_Entry	patient transfer into the OR
2	Sterile_Draping	start of sterile preparation and draping (see playbook)
3	Microsurgery	after the team time-out
4	Emergence	undraping

Frames before the first and after the last segment are unlabeled.

Speaker activity

One label per speech segment and role (Surgeon, OTA, Runner, Anesthesiologist, Nurse, Spectator), as RTTM (file ID runNN) and CSV. Overlapping speech appears as overlapping segments.

Audited (Runs 3, 4, 7): manually checked and corrected. Use these for evaluation.
Pseudo (Runs 1–9): produced automatically from the wearable microphones (300–3000 Hz bandpass, Silero VAD, energy-based attribution between microphones). The CSV adds the attribution contrast in dB per segment. Scored against the audited runs, they reach a diarization error rate of 39.5 / 49.2 / 57.6 % (Runs 3 / 4 / 7, 250 ms collar).

Known issues

Failed wearable microphones: Run 1 (Surgeon, Anesthesiologist), Run 3 (Anesthesiologist), Run 9 (Surgeon). These speakers have no pseudo-labels in those runs, and their speech can show up under the nearest working microphone. The audited Run 3 labels do contain the Anesthesiologist, recovered manually.
Microphone bleed: in Run 7 the Surgeon's speech is mostly attributed to the OTA channel; the pseudo-labels of Run 8 show the same pattern.
Supervision: in several runs (e.g. Run 2), the supervising ophthalmic surgeon was in the OR and is visible in the recordings.
Run 1, ambient microphone: the microscope recording (and with it the ambient microphone) was restarted, so the room track has a gap of about 2.4 min (≈1156–1301 s of the audio timeline).
Run 1 has two microscope recordings (the recording was restarted).
Clocks: the director's log (metadata/events.csv) is 1 h ahead of the recording PCs (folder names).
Visitors: Runs 3 and 4 (see metadata/events.csv); Run 8 has a spectator with a microphone.
Camera 2 is the farthest from the table and partly blocked by the microscope; its depth shows an offset of about 5–10 cm in fused point clouds.
Markers were occasionally lost or occluded (see metadata/events.csv).

Baselines

See the paper for details. Phase recognition (TeCNO, leave-one-run-out over Runs 1–8): best single camera 0.900 frame accuracy; confidence-gated selection of three views 0.898 with a segmental Edit score of 66.5 instead of 39.6. Far-field speaker diarization with an off-the-shelf pyannote.audio pipeline on the ambient microphone: 73.1 % mean DER on the audited runs.

Note for reproducing the phase baselines: the paper's code assigns labels to frames with one constant rate per run (number of frames / annotated duration, about 7.7 frames/s), which places phase boundaries up to 3.3 % of the elapsed time late. For new work, map frames to seconds with sync/room_frame_times/runNN.csv.

Ethics and consent

The study was reviewed by the ethics committee of the Technical University of Munich, which classified it as a technical feasibility study and issued a waiver (2026-333-W-SB). No patients were involved. All recorded persons gave written informed consent, which covers the publication of identifiable video and audio in open-access repositories for scientific research. Faces and voices are therefore not anonymized. Please respect the terms of use.

License

CC BY-NC 4.0 (see LICENSE.md), with gated access for non-commercial research.

Citation

@article{henriques2026bifocalor,
  title   = {Bifocal-OR: A Dual-Scale Multi-Sensor Dataset for Ophthalmic Surgery},
  author  = {Henriques, Angelo and Ross, Andrea and Zhang, Ling and Weisser, Mario and
             Arbabi, Danial and Paulmann, Henri and Sacco, Mateus and Friedrich, Jana and
             Nasseri, M. Ali},
  journal = {TODO},
  year    = {2026}
}

Contact

Angelo Henriques, Technical University of Munich (angelo.henriques@tum.de)

Downloads last month: 348

Total file size:

520 GB