A metric twin of the room, from cameras alone

Proximity is the physical question underneath most sterile-technique rules. Did that sleeve cross the field. How close did the unscrubbed observer get to the instrument table. How far was the door from the patient when it opened.

You cannot answer any of those in pixels. They are metric questions, and they need a metric room.

The gap in existing work

The published options split three ways. You can have geometry without semantics: a reconstruction that knows where surfaces are and nothing about what they are. You can have semantics without 3D: per-frame labels with no metric structure behind them. Or you can have general-scene reconstruction that is not metric, which answers “further” and “nearer” but never “how many centimetres”.

The combination is what makes proximity analytics possible, and it is the claim this line is built on: a metric 3D twin from multi-view RGB alone, no depth sensor at inference, in which every point carries an entity label: staff role or equipment class.

Validation

Per-entity-pair distances are checked against time-of-flight derived ground truth. That is the honest test for a metric claim: not whether the reconstruction looks right, but whether the distances it reports match distances measured by a sensor that was not used to build it. The reconstruction on this site is one second of a real procedure taken straight out of that work.

Why cameras only

Depth hardware in a working theatre is a procurement problem, a sterilisation problem and a mounting problem. Cameras are already a solved install. A method that needs depth at inference is a method that mostly does not get deployed, however well it scores.

Depth is still useful during development, as supervision and as ground truth. The requirement is only that it is absent at inference, which is the moment that decides whether a system can live in the room.