Why several views help
A calibration board can be measured once and photographed at several unknown orientations. The points on it have known planar coordinates, but the rotation and translation in each photograph are not known. We want one intrinsic matrix shared by all views, not a separate unconstrained homography for each image. The board's right angles and equal coordinate scales supply metric information that an arbitrary planar texture alone does not supply.
Assume a rigid planar board with known Cartesian coordinates in a common length unit; each image observes at least four board points including a quadruple with no three collinear, in a configuration that identifies its invertible homography. Assume a fixed pinhole camera with constant invertible intrinsics and initially no lens distortion. Board points must project finitely, and the visible board must not pass through the camera center. The homography-estimation note explains how measured points determine a homography and why normalization and robust filtering matter. Here we start from those homographies and derive the further constraints. A practical estimate is only as reliable as the correspondence, board and camera assumptions behind it.
Two equations per plane orientation
Definition 1.33.1
(Plane-to-image homography with intrinsic matrix)
Choose board coordinates so a point is . Let camera coordinates be , where . With the usual upper-triangular intrinsic matrix
the plane-to-image homography has columns
Definition 1.33.2
(Image of the absolute conic and its two constraints)
Define the symmetric positive-definite image-of-the-absolute-conic matrix . Since the first two columns of a rotation are orthogonal unit vectors, their common unknown homography scale cancels, giving
Definition 1.33.3
(Parameter vector and the bilinear map)
These are homogeneous linear equations in the six independent entries of . Write . For vectors , define
Then . For each view append the two rows and to a matrix . Exact data satisfy . A homography's arbitrary nonzero scale multiplies both rows by its square, preserving the exact equations; under noise this choice changes row weights, so report the normalization and do not interpret the algebraic fit as an invariant geometric maximum-likelihood estimate.
Concept lensGeometric
The constraints as orthogonality in normalized coordinates
Because , we have for all : is the Gram matrix of the ordinary dot product after converting pixel vectors to normalized camera coordinates by . For an exact homography the vectors and are scaled orthonormal vectors, so the two constraints say precisely that they are orthogonal and of equal length. Calibration seeks the inner product under which the board's two axes look orthonormal.
Theorem 1.33.4
(When the linear intrinsic constraints identify a camera)
Suppose the exact data come from one camera of the stated form, and the stacked matrix has rank five. Then its nullspace determines up to nonzero scale. Select its positive-definite sign and normalize the recovered upper-triangular intrinsic matrix to have bottom-right entry one. The result is the unique with positive diagonal and the stated form that is consistent with these constraints. At least three views are necessary for rank five in the unrestricted five-intrinsic-parameter model, but three views are not automatically sufficient.
Proof
(Linear constraints that identify the camera)
Exact data satisfy , where is the nonzero vector of the true . Since has six columns and rank five, its nullspace is the line spanned by , so every solution is with . The matrix is positive definite exactly when , which fixes the sign. Write with upper triangular with positive diagonal. Since is such a factor, uniqueness of the Cholesky factor gives . Hence , and dividing by its bottom-right entry returns . Any camera of the stated form with positive diagonal that is consistent with the constraints therefore has this same normalized .
Proof X-ray
Why a unique conic determines unique intrinsics
The homogeneous equations determine only a ray of matrices, so positive definiteness first chooses its sign. Cholesky uniqueness then identifies the upper-triangular inverse up to a positive scalar. Fixing the bottom-right entry removes precisely that scalar. The argument relies on invertibility and positive focal conventions; an arbitrary rank-two or indefinite matrix cannot be interpreted as this camera.
Each view supplies at most two independent rows; therefore one or two cannot give rank five. Three or more suffice only when the actual stacked matrix reaches rank five. With additional justified intrinsic assumptions, such as known zero skew, fewer views can suffice for a different constrained problem. For noisy data use a smallest right singular vector of as an algebraic candidate and inspect its singular values. Rank five is an exact identifiability condition, while a small separation of the two smallest singular values warns of weak conditioning. Noise can yield an indefinite candidate: neither a sign change nor Cholesky recovery then gives a valid camera without additional modeling or constrained estimation.
Recover intrinsics and pose
The rank condition gives a nonzero symmetric only up to scale. If one sign is positive definite, choose it and compute the unique upper-triangular Cholesky factor with positive diagonal such that . Since is also upper triangular with positive diagonal, uniqueness gives for the positive scale of the estimated . Thus set and divide all its entries by to make the lower-right entry one. This normalization removes the arbitrary scale; it does not impose zero skew or equal focal lengths.
Proposition 1.33.5
(Pose from an exact homography and known intrinsics)
Let hold exactly, with , invertible and , and put , , . Then , , and
for , which is one of the two values . Both values give a proper rotation, and they negate every camera-frame depth of the board points.
Proof
(Reading the pose off the normalized columns)
Since for and , we get , and . Orthonormality of gives and , so and the displayed formulas hold. Replacing by negates and but leaves unchanged, so stays in . A board point has camera depth , which changes sign with .
The sign is not supplied by the homogeneous homography: select the sign for which the observed board points have positive camera depth, if such a sign exists for all of them. For inconsistent noisy data, are not exactly orthogonal or equally long; normalizing one column alone does not create a rotation. Use a declared scale convention and project an approximate frame onto or jointly refine it. In the latter case enforce positive depths for the points used and minimize a specified reprojection objective. Neither step guarantees a global optimum.
Common mistake
(Normalizing the columns separately does not give a rotation)
For noisy data, take and . Normalizing each separately gives and , but , so is not orthogonal. Project the approximate frame onto or refine jointly.
A complete three-view calculation
Worked example 1.33.6
(Recover an unknown camera and one pose)
Use a board with coordinates in , and suppose the exact homographies are represented by
The first view gives and . View 2’s orthogonality equation then gives , and view 3’s gives . View 2’s equal-length equation, after substituting and , reduces to . These five independent constraints imply
The third view's remaining equal-length equation is also satisfied, so it is a consistency check rather than a sixth independent constraint. For , and ; choose to obtain and . Every board point has depth . The other sign puts every one behind the camera. This example uses exact undistorted homographies; it does not claim that a real three-image experiment is well conditioned.
Worked example 1.33.7
(Cholesky recovery with nonzero skew)
Suppose the nullspace vector is proportional to , that is, . This matrix is positive definite: its leading minors are , and . Its Cholesky factor is , because , , and ; indeed reproduces the matrix. Then , and dividing by the bottom-right entry gives . Replacing by gives the same , since the normalization removes the scale. The skew is recovered, not assumed: the normalization imposes neither zero skew nor equal focal lengths.
Common mistake
(Taking the sign of the homography as the sign of the pose)
and represent the same homography. In the three-view example, choosing gives and , so every board point has depth : the whole board lies behind the camera. The sign must be chosen by requiring positive depths.
Counterexample mode
Many frontoparallel views cannot replace independent orientations
For a frontoparallel board let and ; translation, in-plane angle and homography scale may vary across photographs. Both constraints involve only the upper-left block of . For any angle they require and , merely two independent restrictions. In particular the equations do not determine ; infinitely many positive-definite choices remain. Thus moving the board closer, farther or sideways without tilting cannot identify all five unrestricted intrinsics, even with infinitely many noiseless views. Tilts can add information, but their number alone does not prove rank five.
A bounded distortion fit
After an initial pinhole calibration, suppose known estimated poses give ideal normalized rays and inverse-intrinsic measured pixels give distorted normalized coordinates . For use the restricted radial model
With the rays and intrinsics held fixed this is linear in ; stack rows and solve least squares only if the design has rank two. Data solely at the optical center give zero rows; a single nonzero radius cannot separate both coefficients because the second column is that radius squared times the first. A spread of radii helps identification but does not guarantee low uncertainty. This staged fit is linear in the coefficients, not jointly linear in unknown pose, intrinsics and distortion. The initial homographies themselves may have been biased by distortion, so a subsequent joint nonlinear reprojection fit, including the distortion model, is normally needed. State the valid field and check that the radial map is injective on it; tangential distortion and rolling shutter are outside this model.
Worked example 1.33.8
(Two radii determine two coefficients in the restricted fit)
Take ideal normalized points and with observed distorted coordinates and . The first nonzero row gives . The raw second row is ; dividing by gives . Solving gives , . These exact synthetic measurements identify the restricted coefficients conditional on the supplied rays and intrinsics; they are not a test of a physical lens.
Practice and solutions
Quick check
(If a six-column constraint matrix has rank four, how large is its nullspace? Can its linear equations alone identify up to scale?)
Solution · Answer
Rank-nullity gives dimension two. No: a two-dimensional nullspace remains, so infinitely many non-proportional candidates fit; a separate constrained model could add information.
Quick check
(For exact and above, recover using the positive-depth sign.)
Solution · Answer
, , and . Positive depth selects , so , , .
Quick check
(Why cannot repeating the same nonzero-radius distortion measurement determine both radial coefficients, even if it is repeated many times?)
Solution · Answer
At one fixed radius the second column is a constant multiple of the first; repeating rows does not increase design rank beyond one.
Quick check
(A recovered symmetric candidate has eigenvalues . Can changing its global sign make it a valid ? Explain.)
Solution · Answer
No. One sign has eigenvalues , the other ; neither is positive definite, as must be.