Multi-camera and calibration
Several cameras observing one scene, with known intrinsics and extrinsics. This page states the capability and its current evidence boundary in the same breath, because the two are not the same thing.
What exists#
Multi-camera and camera calibration are implemented in the current product surface: the compiler, the API and the runtime all carry them, and intrinsics can be exported with the frame.
What that buys you:
- Stereo and multi-view geometry from known poses, with no calibration rig and no estimation step.
- Cross-camera consistency, because every camera observes the same deterministic scene at the same instant.
- Reprojection you can check: the same actor position projected into two cameras should land in both, and it either does or it does not.
Where the evidence stops#
The latest dedicated visual QA run did not exercise a full multi-camera calibration session. The capability is implemented; the visual proof for it is narrower than the proof for single-camera capture, and this documentation is not going to imply otherwise.
The practical consequence: if a multi-camera dataset is central to your problem, treat it as the thing to validate first in a pilot rather than as a settled part of the product. The visual QA page records what the current evidence does cover.
Coordinate conventions#
The details that cause most of the trouble in multi-view work, stated once:
- Positions are in metres, not Unreal's centimetres.
- The camera frame is named in the capture manifest rather than assumed.
- Quaternions are xyzw ordered, not wxyz.
- Field of view is recorded per frame, so a run that varies it stays interpretable.
All of it is in the per-frame records inside any capture, which is the recommended way to check a convention: read a real file rather than a sentence about one.