Calousel: Extrinsic Calibration of Non-overlapping Multi-camera Systems from Pure Rotation

Calousel calibrates non-overlapping multi-camera systems by rotating the platform on a turntable, allowing all cameras to sequentially observe a single static calibration board.

Abstract

We present Calousel, a practical extrinsic calibration framework for non-overlapping multi-camera systems. Calousel uses pure rotational motion from a single-axis turntable and requires only a single static calibration board. The key idea is to let all cameras sequentially observe the same target under a shared geometric reference, even without overlapping FOV. We integrate these time-separated observations using a latent turntable frame and a 3D error on SE(3) within a global optimization framework. We validate Calousel on both a controlled camera rig and a full-scale vehicle platform with heterogeneous cameras, demonstrating stable performance under realistic sensing conditions and turntable imperfections.

At a Glance

No Overlapping FOV Required

Calousel does not rely on direct inter-camera correspondences.

Single Static Calibration Board

All cameras sequentially observe the same calibration board.

Pure Rotation

A single-axis turntable induces pure rotational motion in a compact setup.

Global SE(3) Optimization

A latent turntable frame integrates time-separated observations.

Why Calousel?

Target-based Methods

Large calibration target setup. — Large target

Multi-target calibration setup. — Multi-target

Large target or multiple target setup
High setup overhead
Pre-measured target poses

Motion-based Methods

Drift error in motion-based calibration. — Drift error

Motion degeneracy in motion-based calibration. — Motion degeneracy

Drift accumulation
Scale ambiguity
Motion degeneracy

Our Method: Calousel

Single static calibration board
Single-axis turntable
Sequential observations
Global SE(3) integration

Method Overview

System Formulation

Calousel models the rotating platform through a latent turntable frame. Although cameras observe the calibration board at different time instants, their observations are connected by the same pure rotational motion.

\[ \boldsymbol{\hat{T}}_{\mathcal{C}_i(t_k)}^{\mathcal{W}} = \boldsymbol{T}_{\mathcal{A}(0)}^{\mathcal{W}} \operatorname{rot}(\hat{z},\;\Omega t_k) \boldsymbol{T}_{\mathcal{C}_i}^{\mathcal{A}} \]

The latent turntable frame separates the time-varying turntable rotation from the time-invariant camera mounting geometry.

Framework

Step 1

Keyframe Selection

Each camera observes the calibration board at different time intervals. We estimate camera-to-board poses and select reliable keyframes with associated covariance.

Step 2

Turntable-Kinematics-Based Initialization

Using the pure-rotation motion model, we initialize the angular velocity, turntable origin, and camera mounting poses from geometric constraints.

Step 3

Global Optimization on SE(3)

We jointly refine all parameters using a latent turntable frame and a 3D pose error on SE(3).

Experiments

We evaluate Calousel on two platforms: a controlled robotic camera rig for quantitative analysis and a full-scale vehicle platform for assessing real-world scalability.

Robotic Camera Rig

The robotic camera rig provides a controlled benchtop setup for quantitative evaluation. It consists of two FLIR Blackfly S cameras mounted on a reconfigurable rig, enabling angular and translational sweeps under non-overlapping FOV configurations. A stepper-motor turntable rotates the rig in front of a single static calibration board.

Robotic camera rig mounting configurations. — Camera mounting configurations

Full-scale Vehicle

We use the full-scale vehicle platform to evaluate Calousel at vehicle scale. The vehicle is equipped with heterogeneous cameras, including FLIR cameras and a RealSense D435i rolling-shutter camera, mounted on a roof rack with non-overlapping views. The platform is rotated on an automotive turntable to validate Calousel under vehicle-scale geometry.

Full-scale vehicle camera configuration variants. — Roof-rack camera mounts

Platform Summary

Item	Robotic camera rig	Full-scale vehicle
# Cameras	2	3
Camera types	2x FLIR Blackfly S	2x FLIR Blackfly S 1x RealSense D435i [RS]
Resolution	1440 x 1080 (FLIR)	1440 x 1080 (FLIR) 1080 x 720 (D435i)
FPS	60	30
Turntable type	Stepper-motor turntable	Automotive turntable
Board config.	4 x 3 r = 0.035 m, d = 0.09 m	5 x 3 r = 0.1 m, d = 0.3 m

[RS] indicates a rolling-shutter sensor.

Results

Camera Rig

RRE < 2.0°

RTE < 2.0 mm

Full-scale Vehicle

RRE 0.9°-1.5°

RTE 5.6-11.7 mm

Overlapping-FOV Check

≤ 0.1° rotation diff.

≤ 0.3 mm translation diff.

Robotic Camera Rig Evaluation

On the controlled camera rig, Calousel maintains stable accuracy across non-overlapping camera configurations, with an average RRE of 1.7° and an average RTE of 1.8 mm.

Rotation components							Translation components
GT			Measured			RRE	Displacement GT			Measured Displacement			RTE
Roll	Pitch	Yaw	Roll	Pitch	Yaw	RRE	Rad.	Tan.	z	Rad.	Tan.	z	RTE
0	0	60	0.7±0.4	0.3±0.2	60.3±0.2	0.9±0.2	-20	0	0	-21.0±1.2	-0.2±1.1	-0.4±0.5	2.0±0.2
0	0	120	0.9±0.2	1.2±0.2	120.0±0.2	1.5±0.3	0	50	0	0.3±1.3	50.5±0.3	0.2±0.6	1.5±0.6
0	0	180	-0.1±0.2	1.8±0.5	179.8±0.5	1.9±0.5	0	0	10	0.8±1.5	-0.2±1.2	9.9±0.5	1.8±1.1

Units are mm for translation and deg for rotation.

Full-scale Vehicle Evaluation

On the full-scale vehicle, Calousel preserves rotation accuracy at vehicle scale. Absolute translation errors increase due to the larger platform geometry, motivating the normalized comparison below.

Cam.	Rotation components							Translation components
	GT			Measured			RRE	Displacement GT			Measured Displacement			RTE
	Roll	Pitch	Yaw	Roll	Pitch	Yaw	RRE	Rad.	Tan.	z	Rad.	Tan.	z	RTE
C₂	0	0	60	0.5	0.7	59.8	0.9	0	0	0	1.1	0.6	4.2	8.3
C₂	0	0	90	0.3	0.1	90.6	1.0	0	150	0	4.0	151.5	3.6	5.6
C₃	0	0	180	-0.2	-0.5	179.1	1.5	0	0	0	1.8	-0.8	2.1	7.7
	-	-	-	-	-	-	-	0	150	0	0.9	145.9	5.2	7.8
	-	-	-	-	-	-	-	0	0	-100	0.6	0.0	-107.8	11.7

Units are mm for translation and deg for rotation. C₂ and C₃ are reported relative to C₁.

Normalized Translation Error Across Scales

To account for platform scale, we normalize each translation error by the corresponding inter-camera distance. Under this normalized metric, the camera rig and full-scale vehicle exhibit comparable component-wise error ratios.

Comparison of normalized translation errors between camera rig and full-scale vehicle. — Comparison of camera rig and full-scale vehicle translation errors using normalized metrics.

Comparison to Target-based Method in Overlapping FOV

In overlapping-FOV stereo configurations, Calousel produces estimates that closely match a conventional target-based calibration method, while retaining the ability to handle non-overlapping configurations.

Setting		Target-based	Ours	Diff. norm
Stereo-A	Rot.	(-0.24, -0.33, 0.61)	(-0.24, -0.33, 0.57)	0.06
Stereo-A	Trans.	(98.76, 1.18, -0.82)	(98.96, 1.27, -0.88)	0.23
Stereo-B	Rot.	(0.11, 0.63, -0.01)	(0.11, 0.63, -0.04)	0.03
Stereo-B	Trans.	(98.30, -10.21, -1.46)	(98.35, -10.21, -1.50)	0.07

Units are mm for translation and deg for rotation.

Reference Accuracies in Prior Work

Because prior methods use different sensor setups and calibration environments, these numbers should be interpreted as contextual references rather than a direct benchmark. Calousel provides competitive accuracy while avoiding large target setups, pre-measured target poses, or additional cameras.

Methods		Rotational error	Translational error
Target-based	Yin et al. (Remote Sens. 2018)	0.06	0.08
Target-based	Liu et al. (Opt. Laser. Eng. 2011)	N/A	0.01-0.08
Motion-based	Xu et al. (RA-L 2022)	2.5	14
Motion-based	Dai et al. (Sensors 2024)	3.7	70.6
Additional-Camera	Robinson et al. (CAIP 2017)	0.02-0.06	2-5
Ours		1.7	1.8

Units are deg for rotation and mm for translation. N/A indicates not reported.

Video

Code

Isornorphism/calousel

Research code for Calousel, including keyframe extraction, turntable-based extrinsic optimization, example data, and evaluation scripts.

https://github.com/Isornorphism/calousel

BibTeX

@INPROCEEDINGS { ghsong-2026-iros,
    AUTHOR = { Gwanhyeong Song and Chaehyeon Song and Ayoung Kim },
    TITLE = { Calousel: Extrinsic Calibration of Non-overlapping Multi-camera Systems from Pure Rotation },
    BOOKTITLE = { IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) },
    YEAR = { 2026 }
}

Acknowledgements

This work is supported by the Korea Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (Grant RS-2023-00250727) through the Korea Floating Infrastructure Research Center at Seoul National University.

This project builds on DiscoCal for accurate circular-pattern intrinsic calibration and board pose extraction. Please cite DiscoCal as appropriate if you use the intrinsic calibration component.