What does your model need to learn?
A model that recognizes surgical phases needs a different dataset from a system that predicts instrument motion. Before counting hours, write down the input, the target, and the setting in which you will test it.
| Research task | Data to specify | A limitation to resolve |
|---|---|---|
| Workflow and phase recognition | Continuous sequences, phase definitions, transition labels, case-level splits | Short clips can remove the context that distinguishes adjacent phases. |
| Instrument and action understanding | Visible tools, temporal labels, and instrument-verb-target definitions | Tool presence alone does not tell you what the tool is doing. |
| Depth and geometry | Paired views, synchronization, calibration, and the delivered per-view raster | Two camera streams do not guarantee usable depth measurements. |
| Robot learning from demonstrations | Video plus the action, state, or other supervision your method requires | Human movement in a headset recording is not a robot command stream. Use the signal planner. |
| Medical device and life sciences research | A defined study question, relevant visual coverage, and separately scoped clinical variables | Patient outcomes and medical records are not automatically included with video. |
Our annotation guide separates phase, event, action, and spatial labels. The research dataset comparison shows how established collections define narrower tasks.
Choose the view before the camera
An endoscopic feed shows the operative field from inside the body. A head-mounted camera records the surgeon’s perspective and can help document visible hand activity, exposed anatomy, and workflow. The useful view depends on the procedure and camera placement. A headset does not substitute for an internal feed when the target is visible only on that feed.
Monocular footage has one view. Stereo has two views whose identity and timing must remain traceable. “2K” is not enough to specify either: state the width and height per view, frame rate, encoding, and whether two eyes are packed into one image. Use the format comparison and storage calculator to make these choices explicit.
Three routes to a collection
| Route | When it helps | What to establish |
|---|---|---|
| Public research datasets | Testing a method against a documented task or baseline | Release version, allowed use, source overlap, and official evaluation protocol. |
| License an existing archive | The relevant procedures have already been recorded | Who can authorize reuse, what footage exists, preparation work, and rights for your intended use. |
| Plan new recording | The required view, specialty, or recording protocol is missing | Institutional permissions, capture feasibility, a pilot, and compensation before collection. |
Operavue discusses both existing archive licensing and new capture partnerships. Surgeons can request an optional recording headset. Minimum guarantees, revenue shares, recording requirements, and acceptance terms are negotiated for each partnership. Explore an archive partnership or the surgeon recording program.
Buy a documented release
Ask for a file manifest, a data dictionary, a label specification, and a record of exclusions alongside the media. Keep the source case identifier stable across clips and derivative files so that one procedure cannot silently appear in both training and test sets. Use project-specific identifiers; naming a file anonymously does not by itself establish de-identification.
The free templates include a synthetic stereo manifest and an acceptance worksheet. They make missing context visible before a large transfer. The local split-audit kit checks source identities in a clip manifest and includes a reproducible synthetic example. For privacy and permission planning, consult the de-identification guide and the relevant institutional team.
Make the pilot answer a decision
A useful pilot ends with a decision such as: “The view captures the target, but we need more complete sequences and a revised ambiguity label.” It should test ingestion, labeling, release documentation, and evaluation setup, as well as image quality. Agree which findings require remediation and which change the scope.
Start with the editable procurement brief. Researchers can submit a buyer inquiry; surgeons and institutions can choose a contribution route. Availability and specifications are established through scoping, not inferred from the examples on this site.