Flexible Camera Calibration using a Collimator System

AI summary

Overview

Research area: Computer vision / photogrammetry — geometric camera calibration.

Technical level: Advanced. The paper builds on planar homography calibration (Zhang's method), the image of the absolute conic, rotation matrices in SO(3), and minimal-solver / hidden-variable techniques.

Scope: The paper proposes a camera calibration framework built around a purpose-built collimator system, deriving an angle invariance constraint and a spherical motion model that reduce the target-to-camera motion from 6DOF to 3DOF, and presents three solvers: a closed-form solver for more than two images, a minimal solver for two images, and a single-image calibration algorithm. It is an extension of the authors' ECCV 2024 conference paper.

What This Paper Is About

Traditional target-based calibration struggles in outdoor, long-range, and large-field-of-view settings, where targets must be enormous, hard to move, and sensitive to unstable lighting; self-calibration is unstable there, and building a full 3D calibration field is operationally complex. Earlier collimator-based approaches solve the target problem but require high-precision goniometers or theodolites to measure collimated ray directions, which is costly and error-prone. This paper asks whether a single designed collimator carrying a star-based pattern on its reticle can serve as a flexible, controllable calibration target whose own optical geometry supplies the missing constraints — including the extreme case of calibrating from just one image.

Key Contributions

1. A designed collimator system for calibration. A known star-based pattern (with an AprilTag at its center for unambiguous localization) is attached to the collimator reticle, providing a reliable, controllable calibration environment suitable for lenses of various focal lengths. The system is portable at approximately 200 mm × 170 mm × 300 mm and requires no angular measurement devices.

2. Angle invariance and the spherical motion model. Exploiting the collimator's ability to produce parallel rays, the authors show that the imaging angle between any pair of reticle points is invariant regardless of camera pose, and prove that the relative motion between calibration target and camera must conform to a spherical motion model — reducing the original 6DOF general motion to a 3DOF pure rotation. A new, simpler proof is offered in this version, validated theoretically and with real collimator images.

3. Two solvers for multiple images plus degeneracy analysis. A closed-form linear solver for N (N > 2) images and a minimal solver for two images are proposed using the spherical motion constraint. The paper additionally presents two degenerate configurations for multiple-image calibration and provides detailed proofs of these cases.

4. Single collimator image calibration. A novel algorithm uses angle invariance to build a cost function that involves only the intrinsics, eliminating the need for camera motion and enabling rapid, batch determination of camera parameters.

Main Findings

• Angle invariance holds by construction: Because rays from any two points on the reticle emerge as parallel beams, the angle between them is identical as observed by cameras at arbitrary poses (θ₁ ≡ θ₂ ≡ θ), independent of camera pose.

• Translation must be zero (Proposition 1): The necessary and sufficient condition for angle invariance is that the translation vector between target and camera is the 0 vector. The proof selects three mutually orthogonal target points A, B, C whose angular distances are all 90°, forms a system of three sphere equations, and finds only two solutions: t = [0, 0, 0]ᵀ and t = [0, 0, −2r]ᵀ. The latter is discarded because the target cannot be positioned behind the camera.

• Motion reduces to 3DOF: The target therefore undergoes pure rotation about the camera optical center with fixed distance r — a spherical motion model where t_cp′ = (0, 0, −r)ᵀ and t_cp = (x, y, −r)ᵀ. Camera and collimator may still move freely; only the relative motion is constrained.

• Constraint counting: There are 8 parameters to estimate — 5 for the intrinsic matrix (f_x, f_y, c_x, c_y, and skew γ) and 3 for t_cp = (x, y, −r)ᵀ. Equation (4.1) supplies 5 independent constraints per image, so a minimum of two images theoretically suffices.

• Closed-form solver derivation: The determinant of the matrix M = [r₁ r₂ −Rt_cp] equals r, a non-zero constant, guaranteeing non-singularity; this yields the scale relation λ_ij = λ_i/λ_j = ∛|H_j⁻¹H_i|. Stacking N images produces a 5N × 11 system D[w; a] = b with the closed-form least-squares solution (DᵀD)⁻¹Dᵀb, after which intrinsics and t_cp are recovered by decomposition.

• Minimal solver for two images: With only two images the coefficient matrix has rank at most 10, so the system is rank-deficient and the true solution is a linear combination of the solution set; the authors instead construct a nonlinear system and use the hidden variable technique, parameterizing the image of the absolute conic Q = K⁻ᵀK⁻¹ and its 6-vector q, deriving per-image constraints (Eqs. 23–27).

• Experimental validation is claimed but numbers are not present in the supplied content: The abstract states the method was evaluated in both synthetic and real-world experiments, verifying feasibility and showing superiority over existing baseline methods. The truncated content provided here does not include the experimental section, so specific datasets, error figures, and benchmark comparisons are not reported in the text available.

• Collimator approach compared favorably to alternatives in design terms: Unlike multi-collimator arrays, which need precise pre-measurement of the whole array and one control point per collimator, and unlike multiple-pinhole or DOE designs that typically demand manual feature extraction, the star-based pattern supports fully automated feature extraction and matching. Earlier work by Hieronymus (2012) is cited as concluding collimator methods can yield accuracy comparable to calibration-field and DOE methods.

Methodology in Plain English

The camera simply looks through the collimator lens at the patterned reticle from several orientations and takes pictures — as easy as photographing a printed board. Because the collimator turns each reticle point into a beam of parallel rays, every reticle point behaves like a point at infinity, so cameras of any focal length can "see" the target from close range and the angle between any two such points never changes as the camera moves.

The authors exploit that fact mathematically. They first prove that if those inter-point angles are to stay fixed, the calibration target cannot translate relative to the camera at all — only rotate. That kills three unknown motion parameters, turning a general six-degree-of-freedom pose problem into a three-degree-of-freedom rotation on a sphere around the camera center. They then combine this spherical-motion constraint with the standard planar homography relationship between the target plane and the image.

From there, the calibration splits into three practical routes. With many images, they write the constraints as one big linear system and solve it in closed form by least squares, then decompose the result to read off focal lengths, principal point, skew, and the camera's position relative to the target. With exactly two images the system is rank-deficient, so they abandon the linear route and set up a nonlinear problem expressed in terms of the image of the absolute conic, solving it with the hidden-variable technique. With just one image, they use angle invariance directly to build a cost function containing only the intrinsics — no camera motion needed — which makes fast, batch calibration possible. Rotation matrices are re-orthogonalized by singular value decomposition, and the paper also identifies two configurations in which multiple-image calibration degenerates.

Why This Matters

Impact on research: The work repositions the collimator from a precision-instrument-dependent lab tool into a portable, geometry-driven calibration target. Introducing angle invariance as a formal constraint, proving the spherical motion model, and providing a minimal two-image solver and a single-image solver gives the calibration community new constraint structures and lower-input calibration options. Published as an extension of an ECCV 2024 paper, it adds a new proof, degeneracy analysis, a single-image algorithm, and additional real-data experiments (spherical motion validation, structure-from-motion results, single-image calibration results, and further discussion).

Real-world applications:

• Long-range depth estimation, where cameras work at large distances and standard printed targets cannot span the field of view.
• Wind turbine blade monitoring from a distance, where on-site target placement is impractical.
• Outdoor and constrained-space deployments with poor or uncontrollable illumination, where the collimator's built-in planar array light source provides a stable environment.
• Drone- or vehicle-mounted camera calibration, where portability and freedom from goniometers or theodolites matter.

Industry relevance: The system is designed to be portable and low in hardware complexity compared to multi-collimator arrays and to direction-measurement setups, and it supports automated feature extraction and matching — which points toward faster, cheaper calibration workflows in field service, aerospace, and industrial inspection. The authors note the design suits calibrating lenses of various focal lengths and can be used in complex environments like outdoor settings.

Future Directions

• Quantify accuracy on standard benchmarks. The supplied content does not report dataset names, image counts, or reprojection-error numbers; establishing how the single-image and two-image solvers compare against Zhang's method and against goniometer-based collimator calibration on standard datasets is the natural next empirical step.
• Characterize the degenerate configurations more broadly. The paper identifies two degeneracies for multiple-image calibration; whether analogous degeneracies exist for the two-image minimal solver and the single-image algorithm is left open.
• Handle lens distortion and non-pinhole cameras. The paper's derivations use the pinhole model with skew; the related-work section notes fisheye and omnidirectional models (Kannala-Brandt, dual-sphere, omnidirectional) that this formulation does not yet address.
• Reduce reliance on machining precision. The authors criticize multiple-pinhole masks for their sensitivity to machining errors; a systematic error-propagation study of the star-based reticle under the proposed algorithms would clarify how much this design actually mitigates that issue.

Target Audience

Researchers and engineers in photogrammetry, 3D vision, and optical metrology who work on camera calibration; practitioners deploying cameras for long-range, large-field-of-view, or illumination-constrained measurement; and readers already familiar with planar-homography calibration who want to understand how collimator optics can supply additional geometric constraints — including single-image calibration. Given the algebra (SO(3) parameterizations, the image of the absolute conic, hidden-variable solvers), the paper is best suited to readers with a solid background in multi-view geometry.

Authors’ abstract

Camera calibration is a crucial step in photogrammetry and 3D vision applications. This paper introduces a novel camera calibration method using a designed collimator system. Our collimator system provides a reliable and controllable calibration environment for the camera. Exploiting the unique optical geometry property of our collimator system, we introduce an angle invariance constraint and further prove that the relative motion between the calibration target and camera conforms to a spherical motion model. This constraint reduces the original 6DOF relative motion between target and camera to a 3DOF pure rotation motion. Using spherical motion constraint, a closed-form linear solver for multiple images and a minimal solver for two images are proposed for camera calibration. Furthermore, we propose a single collimator image calibration algorithm based on the angle invariance constraint. This algorithm eliminates the requirement for camera motion, providing a novel solution for flexible and fast calibration. The performance of our method is evaluated in both synthetic and real-world experiments, which verify the feasibility of calibration using the collimator system and demonstrate that our method is superior to existing baseline methods. Demo code is available at https://github.com/LiangSK98/CollimatorCalibration