Rolling shutter

A rolling-shutter camera exposes rows sequentially, so one image spans a range of times. CamCalib solves the readout time as a parameter rather than correcting for it afterwards.

Enable it

config.yaml

cameras:
  - camera_id: 0
    shutter_type: rolling      # global | rolling

Per camera — a rig can mix global- and rolling-shutter cameras freely, and the released 4-IMU/4-camera rig does exactly that.

In a camera–IMU config, readout can be estimated in the cam-IMU stage instead:

cam_imu.yaml

camera_input:
  estimate_readout: true

Why it is solved jointly

Projection, time offset, rolling shutter, extrinsic composition, and SLERP pose interpolation live in one unified factor. A row’s observation is projected against the trajectory at that row’s time, not the frame’s — so readout is constrained by the same residuals as everything else instead of being a post-hoc correction with no derivative.

What to expect

Readout is among the weakest parameters in the graph. It is only excited by motion fast enough that rows see meaningfully different geometry — a slow, smooth recording will not constrain it, and the estimate will drift toward whatever the initialisation suggested.

Two checks worth more than a single run’s residual:

• Repeatability. Calibrate several sequences of the same rig with batch mode and compare. A readout that moves between sequences was not observed in them.
• Round-trip. camcalib roundtrip catches a readout that absorbed error belonging to the time offset.

Readout and time offset trade against each other

Both shift when an observation is considered to have happened. With insufficient motion the solver cannot separate them, and the pair can land anywhere along a ridge while the residual stays flat. If a rolling-shutter camera’s td looks displaced relative to its global-shutter neighbours, suspect this before the hardware.

Reporting

Readout appears in the HTML report’s temporal-calibration section, merged with the time offset — they are read together for the reason above. Regression comparison reports break out per-collection readout and time offset so drift across a dataset family is visible.

The residual and its Jacobians: Camera model.