In geometric computer vision, cameras are traditionally modeled using the elegant pinhole abstraction. A core tenet of this standard model is radial symmetry: the assumption that the camera's optical axis pierces the image sensor at its exact geometric center $(W/2, H/2)$. This point serves a dual purpose. It is treated both as the principal point (the origin of the camera coordinate system on the image plane) and the center of distortion (CoD) around which lens warping occurs.
As Mendonça et al. point out, fixing the principal point at the geometric center provides crucial numerical stability during optimization. Without this constraint, self-calibration algorithms easily diverge because focal length and principal point location are highly coupled. Consequently, standard toolkits and solvers traditionally treat the optical center as a known, perfectly centered constant.
Why the Centered Assumption Fails
In the real world, the assumption that the optical center aligns perfectly with the image center is routinely violated by two distinct forces: physical manufacturing imperfections and digital post-processing.
Physical Misalignments
Lenses are complex mechanical assemblies. As Sasián et al. explain, even if a lens barrel is perfectly centered mechanically, internal manufacturing flaws like "wedge errors" within individual glass elements can cause the optical axis to deviate from the mechanical axis. Furthermore, in specialized hardware like non-frontal cameras, the sensor plane may be intentionally or accidentally tilted, shifting the entrance pupil relative to the lens Kumar 0001 et al.. In ultra-miniature hardware like micro-cameras, manufacturing tolerances are large relative to the tiny sensor size, causing the center of radial distortion to routinely drift away from the principal point by significant margins Dias.
Digital Manipulation
Even if a camera is perfectly manufactured, modern image pipelines break the centered assumption. When an image is asymmetrically cropped or digitally zoomed, the original optical center is shifted away from the new geometric center of the resulting image file Larsson et al., Brito et al..
The Geometric Cost of Misalignment
What happens when an algorithm assumes the center of distortion is at the image center, but it is actually offset by a few pixels?
Standard radial distortion models apply a correction factor based on the radial distance $r$ from the center of distortion:
$$x_{\text{corrected}} = x_{\text{distorted}}(1 + k_1 r^2 + k_2 r^4 + \dots)$$
If the origin of $r$ is misplaced, the correction vectors point in the wrong direction. This miscalculation transforms a simple, symmetric radial warp into a complex, asymmetric distortion pattern.
Ideal Centered Model: Real-World Misaligned Model:
+-----------------+ +-----------------+
| \ | / | | \ | / |
| -- CoD -- | | -- CoD -- |
| / | \ | | / | \ |
| | | |
+-----------------+ +-----------------+
(CoD at Geometric Center) (CoD Offset by Manufacturing/Crop)
The consequences for downstream geometric tasks are severe. Larsson et al. demonstrated that ignoring this principal point offset in camera pose estimation results in catastrophic errors, inflating rotation errors to over 14 degrees in cropped images. Similarly, Kocur et al. showed that minor deviations in the principal point degrade the accuracy of focal length estimation from fundamental matrices, as the model lacks the flexibility to absorb the geometric noise caused by the offset.
Algorithmic Strategies for Finding the True Center
To overcome these limitations, researchers have developed specialized mathematical frameworks that treat the optical center as an explicit unknown variable rather than a fixed anchor.
Minimal Solvers
Instead of relying on heavy iterative optimization that might diverge, Larsson et al. introduced "exactly minimal" solvers that solve for camera pose alongside an unknown principal point by deriving novel polynomial constraints. Similarly, Bujnak et al. and Kuang et al. formulated minimal solvers to handle unknown radial distortion centers in absolute and relative pose estimation, ensuring that the geometry remains solvable even when dealing with uncalibrated or cropped hardware.
Epipolar Invariants
A particularly elegant insight comes from Brito et al., who observed that while radial distortion warps most straight lines into curves, any 3D line that passes directly through the camera's optical center projects as a perfectly straight line in the image. By identifying these unique straight epipolar lines from raw image correspondences, their method computes a closed-form solution for the center of distortion without needing any prior scene structure or calibration patterns.
1D Radial Invariance
Another approach leverages the "1D radial camera model." In this view, we ignore the magnitude of radial distortion entirely and focus solely on the angular direction of points radiating from the center of distortion. Because this angular direction remains invariant under radial warping, researchers can perform 3D reconstruction and calibration by isolating the distortion center from the radial profile itself Multi-View Geometry of 1D Radial Cameras and its Application to Omnidirectional Camera Calibration, Non-parametric Structure-Based Calibration of Radially Symmetric Cameras.
Moving Beyond Parametric Symmetry
As lenses become wider and manufacturing configurations more complex, even estimating a single off-center point is sometimes insufficient. High-performance wide-angle lenses often exhibit "mustache distortion"—a non-monotonic, asymmetric warping pattern. As Lin et al. demonstrate, models that assume perfectly circular radial symmetry fail to correct these complex profiles because they cannot map non-uniform distortions.
To address this, the field is moving toward non-parametric camera models. For instance, Wang et al. introduced an incremental Structure-from-Motion pipeline that replaces rigid parametric models with flexible, spline-based representations. By dynamically estimating the local projection behaviors across the sensor rather than enforcing a global radial center, these approaches can self-calibrate almost any arbitrary lens configuration. Similarly, Porfiri Dal Cin et al. decoupled calibration from specific lens models by learning a model-agnostic camera representation, effectively neutralizing the systematic errors introduced by misaligned principal points and lens skew.
Key Foundational Work
Precise Omnidirectional Camera Calibration (2001) — contributed 1 concept elements to this direction.
A Simple Technique for Self-Calibration (1999) — contributed 1 concept elements to this direction.
Camera Pose and Calibration from 4 or 5 Known 3D Points (1999) — contributed 1 concept elements to this direction.