Motivation
Two photographs of a three-dimensional scene need not be related by a single homography. Nevertheless, if a point is identified in both, its second image lies on a predictable line. This weaker relation is useful: it narrows correspondence search and constrains relative camera motion without requiring every point to lie on one plane. The relation is a constraint on possible matches, not a complete reconstruction by itself.
We assume ideal central pinhole cameras and finite, undistorted image points. The earlier note gives projection; the robust-alignment note supplies normalized homogeneous least squares and warns about outliers. Here we derive the two-view equations and identify the extra assumptions needed to interpret them as physical motion.
Geometry and the linear estimator
Definition 1.35.1
(Calibrated two-view setup and the essential matrix)
In camera 1 coordinates let a scene point be , where is a calibrated ray and . Camera 2 coordinates are , with and . For the skew matrix satisfying , the essential matrix is
Proposition 1.35.2
(Epipolar constraint)
Let be the second calibrated ray, where has nonzero finite depth in camera 2, so that for some . Then
Proof
(A scalar triple product vanishes)
Since and ,
Both terms vanish, because is orthogonal to and to . Equivalently, , and lie in one plane, so their scalar triple product is zero.
Proof X-ray
The constraint is necessary before it is diagnostic
A correct pair satisfies the epipolar equation, but every point on the epipolar line also satisfies it. The scalar equation alone cannot say which point is the true match, nor establish positive depth. The rank condition identifies a matrix only under its own nondegeneracy hypothesis; physical recovery adds calibration, essential-matrix and visibility conditions.
The vector is a line in the second normalized image when nonzero; points on it satisfy the equation. At the first-image epipole , , and an equation with zero line gives no useful restriction.
Definition 1.35.3
(Fundamental matrix)
For pixel coordinates , with invertible known intrinsics, set
With , , and the entries of stacked row by row in , each correspondence contributes the row
Stacking the rows gives . If exact data give rank eight, the nullspace is one-dimensional, so the nonzero matrix is determined up to scale. Eight rows alone do not ensure rank eight. With noise solve using a smallest right singular vector; check whether a small spectral gap makes the answer unstable. This minimizes an algebraic error, not geometric image displacement, and the resulting matrix need not have rank two.
Translate and isotropically scale each image's measured points separately to zero centroid and RMS radius , provided that radius is positive. For transformations , fit in normalized coordinates, then return . This follows by substitution into . Normalization improves numerical scaling but cannot create rank missing from the scene. With , , replacing by zero gives a nearest rank-at-most-two matrix in Frobenius norm in those normalized coordinates. This is a constrained algebraic estimate, not necessarily a nearest geometric fit. A subsequent robust, geometric refinement can use the original measurements; it needs a separately stated residual and outlier model.
Worked example 1.35.4
(One row of the linear system)
Let , so , and take , . Then , , , and the row is , giving . Directly, and . The two computations agree because merely lists the nine products appearing in . The nonzero value shows that this pair is not an exact match for this ; a correct match would give .
Worked example 1.35.5
(Normalizing four points)
Take the points . The centroid is and each point is at squared distance from it, so the RMS radius is . The scale gives
which sends the points to , each at distance from the origin. Edge case: if all measured points coincide, the RMS radius is zero, no finite scale is defined, and normalization is impossible; the matches carry no information about in that case.
Common mistake
(Skipping normalization because the algebra is exact)
In pixel coordinates of order the columns of containing and are of order while the last column is . The singular value decomposition of such a matrix is dominated by the large columns, so measurement noise perturbs the smallest singular vector much more than the exact algebra suggests. Normalization makes the columns comparable. It is a numerical device that changes neither the model nor the rank of exact data.
The calibrated constraint and gauge
Theorem 1.35.6
(Singular values of an essential matrix)
For and nonzero , the singular values of are . Conversely every real nonzero matrix with singular values , , is an essential matrix up to nonzero scale.
Proof
(Skew-matrix identity and oriented singular bases)
Forward direction. Since , we have . The matrix has eigenvalue on the plane orthogonal to and eigenvalue on the line spanned by , and conjugation by the orthogonal matrix does not change eigenvalues. Hence the singular values of are .
Converse. Write with orthogonal. Because the third singular value is zero, the third columns of and can be negated without changing ; do so until . The canonical matrix satisfies , where is the rotation through degrees around . For we have , so
With and this is .
Thus equal nonzero singular values, not rank two alone, characterize the calibrated model. For a noisy tentative , its Frobenius projection onto this set uses for both nonzero singular values and zero for the third. This operation depends on the chosen calibrated coordinate metric; projecting to rank two is a different operation.
Definition 1.35.7
(Gauge freedom)
Let be camera matrices and scene points whose projections are the observations. A gauge freedom is a change of the unknowns that leaves every ideal observation exactly unchanged. For any invertible matrix , replacing by and by is one, because .
Decomposing an essential matrix has several algebraic motion candidates. With corresponding rays, triangulation and positive-depth checks in both views can select a physically admissible candidate under nondegenerate conditions; they do not remove every ambiguity or cure wrong matches. The magnitude of cannot be recovered from two unscaled images: multiplying every scene coordinate and by the same positive constant leaves each projection unchanged. The first camera frame fixes a world origin and orientation, but not this scale. If neither intrinsics nor metric world structure is known, an invertible projective change of world coordinates, applied inversely to the camera matrices, also preserves the images. Metric interpretation therefore requires further calibration or scene constraints.
Worked example and degeneracy
Worked example 1.35.8
(A sideways translation)
Let , , , and . Then , , and . Here
The second-image line is , which contains . It contains infinitely many other points, including : the epipolar test alone accepts them too. Scaling both and by two gives the same two normalized image points.
Counterexample mode
A perfect planar fit need not identify two-view motion
Suppose all matches obey for an invertible plane homography . For any nonzero vector , define . Then because is parallel to and . Nonparallel choices of yield nonproportional admissible matrices because is invertible, so even arbitrarily many exact planar matches cannot by themselves select a unique fundamental matrix. Likewise, pure rotation has and gives a homography; the essential matrix derived from the actual motion is zero and provides no nonzero baseline for triangulation.
A real lens can invalidate the straight epipolar-line model in raw pixels. For example, in normalized camera coordinates a restricted radial model is , , . Its center, parameters and valid field must be calibrated; an inverse can be ambiguous where the radial map is not one-to-one. Tangential distortion, rolling shutter, finite aperture and moving objects are outside this model. Estimating directly from distorted pixels generally approximates curved constraints, so a low residual on a small region does not validate the model over the whole image.
Concept lensStructural
Coordinate freedom is not measurement noise
Gauge describes transformations that leave ideal observations exactly unchanged. Noise makes measured constraints inconsistent. Anchoring a coordinate frame removes a chosen representation freedom; it cannot manufacture baseline, nonplanar depth variation or correct correspondences.
Practice and solutions
Quick check
(For the sideways-translation example, test . Can it be a correct ideal match?)
Evaluate with the computed there.
Solution · Answer
Its epipolar value is , so it fails the necessary constraint and cannot be an ideal match under this motion.
Quick check
(If has rank seven, does eight or more measured pairs imply a unique homogeneous estimate? Explain.)
Count the dimension of the nullspace of .
Solution · Answer
No: rank-nullity gives a two-dimensional nullspace in nine unknowns; a count of rows does not determine rank.
Quick check
(A nonzero matrix has singular values . Is it an essential matrix up to scale? What is its Frobenius projection onto the essential singular-value pattern?)
Compare the two positive singular values, then average them.
Solution · Answer
No: its two positive singular values differ, and scaling preserves their ratio. Keep the singular vectors and replace the values by .
Quick check
(The radial map is for . Why must a purported inverse specify a domain?)
Find where increases and where it decreases.
Solution · Answer
rises then falls: and both map to zero. Restricting to an injective interval, such as , is necessary for a single-valued inverse there.