Multi-camera / multi-IMU rigs

Multi-sensor rigs are the case CamCalib is built for. There is no separate stereo step and no pairwise composition: every camera and every IMU enters one factor graph, so correlations between them are kept rather than discarded.

Several cameras

Add entries and choose the reference:

config.yaml

cameras:
  - camera_id: 0
    source_type: rosbag_topic
    source_path: /data/rig.bag
    topic: /cam0/image_raw
    distortion_model: equidistant
  - camera_id: 1
    source_type: rosbag_topic
    source_path: /data/rig.bag
    topic: /cam1/image_raw
    distortion_model: equidistant
  - camera_id: 2
    source_type: rosbag_topic
    source_path: /data/rig.bag
    topic: /cam2/image_raw
    distortion_model: radtan       # models may be mixed
    shutter_type: rolling

reference_camera: 0
optimize_time_offset: true         # for unsynchronised cameras

Extrinsics are expressed against reference_camera. Camera-to-camera transforms follow by composition through the graph, which is what keeps a five-camera rig consistent.

Cross-FOV rigs — a fisheye and a narrow lens on the same rig — are where time-offset estimation usually fails. CamCalib runs two independent estimators (PnP-based and M5 homography cross-correlation) precisely for this case.

Several IMUs

The first entry in imu_topics is the base IMU: the reference clock, and the only IMU that carries the trajectory. Everything else is auxiliary, related by a rigid transform and a time offset.

cam_imu.yaml

input:
  imu_topics: ["/imu0", "/imu1", "/t265/imu"]
  imus:
    - {}                       # base — needs the accel
    - {}
    - {gyro_only: true}        # aux with an untrustworthy accel

Choose the best sensor as base — highest rate, lowest noise. Its errors propagate to the whole rig.

gyro_only is per-IMU and not permitted on the base. What it does and when to prefer it over estimate_t_ga: IMU model.

Why redundancy helps

Several IMUs observe the same motion. That redundancy constrains parameters a single IMU leaves weak — time offsets, readout, and IMU intrinsics — which is the reason to carry the extra sensors at all. It is also what makes the round-trip check meaningful on these rigs: a wrong calibration usually cannot satisfy every IMU at once.

Sensor numbering when comparing with Kalibr

CamCalib numbers sensors by role; Kalibr numbers by input order. Getting this wrong silently compares the wrong pair. For the released rigs:

4 IMUs + 3 cameras

Sensor	BlackFly	T265 L	T265 R	GX3-25	GX3-35	Xsens MTI-100	T265 IMU
Kalibr	cam0	cam1	cam2	imu0	imu1	imu2	imu3
MVIS	Cam2	Cam0	Cam1	Ib	Is0	Is1	Is2

4 IMUs + 4 cameras

Sensor	T265 L	T265 R	ELP L	ELP R	GX3-25	GX3-35	Xsens	T265 IMU
Kalibr (T265)	cam0	cam1	—	—	imu0	imu1	imu2	imu3
Kalibr (ELP)	—	—	cam0	cam1	imu0	imu1	imu2	imu3
MVIS	Cam0	Cam1	Cam2	Cam3	Ib	Is0	Is1	Is2

CamCalib reads and writes Kalibr camchain and camchain-imucam files directly — see Result formats.

Worked runs

4 IMUs + 3 cams · 4 IMUs + 4 cams