Calibration and spatial alignment

Training on multi-camera data assumes you know where each camera is. Hand-tuned extrinsics drift as mounts get bumped, and the model sees the error as noise. Rayform keeps calibration as data: measured, versioned, checked and attached to every run.

Calibrate

Static cameras are calibrated against a ChArUco board; wrist cameras with a hand-eye routine while the arm moves through a set of poses.

rayform calibrate intrinsics --camera front
rayform calibrate extrinsics --rig rig.toml --board charuco_7x5_40mm
rayform calibrate hand-eye --rig rig.toml --camera wrist_left --poses 20

Each command writes a new calibration version and prints its residual error:

front      → base_link   reprojection 0.31 px · 3D residual 1.6 mm
wrist_left → tool0       reprojection 0.42 px · 3D residual 2.1 mm

Accuracy depends on your cameras, board and setup; the residual is reported so you can decide whether it is good enough rather than assume it.

Check for drift

rayform calibrate verify --rig rig.toml

Verification compares the current views with the stored calibration and flags cameras that have moved. rayform check also warns when a run's depth disagrees with the calibration more than usual.

Using calibration

• Every run stores the calibration version it was captured with.
• rayform export --frame base writes camera poses and, on request, point clouds in the base frame.
• Recalibrating never rewrites old runs; export a run again to apply a newer calibration.