Motivation / Problem Setup

For calibrated two-view geometry,

$$ x_2^\top E x_1=0,\qquad E=[t]_\times R. $$

A generic rotation matrix $R\in SO(3)$ has nine entries constrained by

$$ R^\top R=I,\qquad \det R=1. $$

The symmetric equation $R^\top R=I$ imposes six independent constraints—three unit-length and three orthogonality constraints—leaving

$$ 9-6=3\ \text{rotation DoF}. $$

Together with two translation-direction DoF, a generic relative pose therefore has

$$ 3_{\text{rotation}}+2_{\text{translation direction}}=5\ \text{DoF}, $$

so five point correspondences are minimal.

If the rotation axis is known, the rotational freedom reduces to only the angle:

$$ 1_{\text{angle}}+2_{\text{translation direction}}=3\ \text{DoF}. $$

Hence three point correspondences are minimal.

<aside> 🎯

Research focus. Existing methods already solve this three-point problem. The goal here is to understand the geometry behind the four solutions, connect the existing formulations, distinguish chart singularities from genuine data degeneracy, and explore globally compatible three-view extensions.

</aside>

Previous Works

Formulation Core equation What it establishes
Generic essential variety $2EE^\top E-\operatorname{tr}(EE^\top)E=0$ Dimension 5, degree 10; basis of the calibrated five-point problem.
Fraundorfer et al. $E=\begin{bmatrix}a&-b&c\\b&a&d\\e&f&0\end{bmatrix}$ Three-point structured-$E$ solver; the essential cubics reduce to a quartic.
Sweeney et al. $(q^2M+qC+K)t=0$ Direct pose solver using the cotangent half-angle and a quadratic eigenvalue problem (QEP).
Guan et al. One affine match gives three linear constraints. Replaces three point matches with one affine correspondence; the known-vertical problem remains quartic.
Choi and Kim $(\theta,\phi)$: yaw and planar translation direction Planar motion has two observable DoF, giving a two-point minimal problem.

Planar-motion two-point methods impose the stronger condition that translation also lies in a plane. They are related but do not describe the same three-DoF family, so they are background rather than the central comparison.

Benchmark

The accompanying constrained-E repository provides a controlled comparison of:

Method Implementation compared Role
Sweeney QEP Cotangent-half-angle QEP, $q=\cot(\theta/2)$ Standalone baseline
Fraundorfer quartic Structured-$E$ nullspace and quartic elimination Standalone baseline
Chart-aware wrapper Tangent QEP $\rightarrow$ cotangent QEP $\rightarrow$ structured-$E$ fallback Wrapper around equivalent formulations, not a new solver family