For clinical video recognition, start with Cholec80, CholecT50, and Endoscapes. For controlled skill tasks, inspect JIGSAWS. ImitateCholec adds ex vivo imitation-learning demonstrations; Open-H-Embodiment combines several healthcare robotics settings. Choose by the target, recorded setting, and available signals, then check the release and terms.
Compare the task, not just the size
This selection spans clinical visual recognition, controlled skills, ex vivo imitation learning, and mixed healthcare robotics. It is a starting point, not an exhaustive catalog or ranking. The review checks primary documentation, not recordings or trained models. Operavue does not own, host, or sell these datasets.
Comparison version 1.1 · Documentation checked October 18, 2026. Sources and release references appear under each entry.
| Collection | Research focus | Recorded setting |
|---|---|---|
| Cholec80 | Phase recognition and tool presence | Clinical laparoscopic procedures |
| CholecT50 | Instrument, verb, and target recognition | Clinical laparoscopic procedures |
| Endoscapes | Scene segmentation, detection, and CVS assessment | Clinical laparoscopic procedures |
| JIGSAWS | Gesture recognition and technical skill assessment | Controlled bench-top training |
| ImitateCholec | Imitation learning for clipping and cutting | Ex vivo porcine tissue on one dVRK Si setup |
| Open-H-Embodiment v1 | Healthcare robot learning across multiple embodiments | Clinical, ex vivo, phantom, and simulated settings |
Cholec80
Video captured at 25 fps; phase labels at 25 fps and tool presence labels at 1 fps.
- Scope and count unit
- 80 laparoscopic cholecystectomy videos
- Access and use
- CC BY-NC-SA 4.0, according to CAMMA. The public license includes a noncommercial restriction.
- Before using it
- A tool-presence label does not identify an instrument’s action or target. Check overlap with related CAMMA datasets before choosing splits.
- Release reference
- Cholec80 release documented on the CAMMA dataset page
- Documentation checked
CholecT50
Action triplets annotated at 1 fps. The release also provides phase labels and bounding boxes for a five-video subset.
- Scope and count unit
- 50 laparoscopic cholecystectomy videos
- Access and use
- CC BY-NC-SA 4.0. Use the official repository for release files, variants, and terms.
- Before using it
- 45 videos come from Cholec80. Treat the collections as related, not as independent training and test sources.
- Release reference
- CholecT50 documented release; challenge variants have separate protocols
- Documentation checked
Endoscapes
The public release has distinct CVS, bounding-box, and segmentation subsets. Segmentation masks cover a subset of 50 videos.
- Scope and count unit
- Selected regions from 201 laparoscopic cholecystectomy videos
- Access and use
- The repository describes noncommercial scientific use under CC BY-NC-SA 4.0.
- Before using it
- Do not assume all frames have every label type. The public segmentation subset differs from the larger private set used in some papers.
- Release reference
- Public CVS, bounding-box, and segmentation subsets in the official repository
- Documentation checked
JIGSAWS
Stereo endoscopic video, robot kinematics, gesture annotations, and skill ratings collected from eight surgeons.
- Scope and count unit
- Suturing, knot-tying, and needle-passing tasks
- Access and use
- The original release page specifies academic research, not commercial use. Follow the owner’s current access instructions.
- Before using it
- These are controlled training tasks. Results on this collection do not establish performance during complete patient operations.
- Release reference
- Original Johns Hopkins CIRL release documentation
- Documentation checked
ImitateCholec
Stereo endoscope images, two wrist-camera views, robot poses and joint states, with optimal and recovery demonstrations.
- Scope and count unit
- 18,343 short demonstrations from 34 tissue samples; not independent patient operations
- Access and use
- The published release metadata identifies CC BY 4.0 for data. The code uses MIT. Current media retrieval was not independently verified.
- Before using it
- One procedural phase, repeated tasks, and no hardware synchronization. The paper does not guarantee exact video/kinematic alignment; some cutting jaw trajectories are interpolated.
- Release reference
- Johns Hopkins data release v1.0, DOI 10.7281/T1PF3FYK
- Documentation checked
Johns Hopkins data release ↗ · Official repository ↗ · Scientific Data descriptor ↗ · Publisher-deposited release metadata ↗
Open-H-Embodiment v1
Paired video and robot kinematics in LeRobot v2.1 format. Views, action definitions, annotations, and additional sensors vary by component.
- Scope and count unit
- A collection spanning surgical robotics, ultrasound, and endoscopy; component coverage varies
- Access and use
- The collection card lists CC BY 4.0. SurgSync’s original card lists CC BY-NC 4.0, while its Open-H component documentation lists CC BY 4.0. Ask the maintainers to clarify this difference before assuming commercial permission.
- Before using it
- Check the chosen component and its timing documentation. This is not a uniform collection of complete patient operations. Current v2 contribution guidance describes a different format from hosted v1.
- Release reference
- Hosted v1, Hugging Face revision e29dda7cabf2a2626634c7822db27695553ae523
- Documentation checked
NVIDIA-hosted release card ↗ · Project and release guidance ↗ · Original SurgSync terms ↗ · Open-H SurgSync component documentation ↗
One source can appear in several collections
A new dataset name does not necessarily mean new source footage. CAMMA documents overlaps between Cholec80, CholecT50, and Endoscapes. Read the official overlap analysis before combining these releases.
If you pretrain on one collection and evaluate on another, a shared source case can undermine a claim that the evaluation uses unseen operations. Record source procedures across every split, including unlabeled pretraining data. For repeated lab demonstrations, preserve the tissue, operator, and session grouping relevant to the experiment. Different clip names do not establish independence.
Our local split-audit tool checks supplied source identities. It cannot discover an undocumented relationship between collections. Establish that mapping before interpreting a clean result.
Does public access permit commercial model training?
The terms differ. The CAMMA entries describe noncommercial use; the original JIGSAWS page specifies academic research. ImitateCholec’s published data metadata identifies CC BY 4.0. Open-H’s collection label does not explain the SurgSync component’s different original-source label. These summaries are not a blanket permission for a proposed use.
Read the applicable release and component terms, attribution requirements, and any separate permissions. A code license does not automatically govern media. Where documentation differs, seek clarification from the maintainers. Operavue accepts surgical data sourcing inquiries without promising access to the third-party collections listed here.
A short selection protocol
- Define the target and test setting: clinical video recognition, controlled skill, or robot learning.
- Identify the released subset, count unit, and label coverage. Match them to the paper’s experimental material.
- For robot learning, inspect observations, state, action definitions, and timing. Use the signal planner to specify what your method requires.
- Review permitted use with the dataset owner and preserve the relevant release documentation.
- Map source relationships, choose evaluation groups, and run a small ingestion check.
Use and cite the comparison
The CSV is generated from the same records as this page. It includes scope, signals, release references, caveats, primary URLs, and the check date. It contains reference metadata, not recordings. Save a dated copy with your project; the linked sources may change.
Operavue. Surgical Video Datasets: A Research Comparison. Version 1.1. October 18, 2026. https://operavue.com/resources/surgical-video-datasets.
When using a dataset in research, cite its original paper and release as the authors request. This comparison does not replace those citations. Version 1.1 adds ImitateCholec, Open-H-Embodiment, individual release references, and the CSV export.
For task definitions, read surgical video annotation. Report corrections through our general inquiry form with a link to the relevant release documentation.
FROM SPECIFICATION TO PARTNERSHIP
Scope the data your team needs.
Discuss existing footage, a custom capture program, or clinical annotation around a specific research task.
Share a non-confidential brief ↗