Motivation
A bright pixel may arise from a highly reflective surface, favorable lighting, or both. One image generally does not separate these causes. With a fixed camera and several known lighting directions, repeated observations of the same visible surface point can constrain its normal and diffuse reflectance. The HDR chapter uses illumination to discuss image formation; this unit works the inverse question under a deliberately narrow model.
A calibrated diffuse model
Definition 8.1.1
(Observations of one fixed point)
Let be an unknown unit surface normal and let be its unknown diffuse coefficient. For images, let each known unit vector point from the surface toward light . Assume the camera and surface do not move, all observations correspond to the same point, and the point is illuminated and visible in every image. With known light strength and exposure corrected, no saturation or specular term, and , the linear-intensity model is
The error collects measurement and model error. Any fixed Lambertian normalization factor is absorbed consistently into the calibrated coefficient; a display-encoded value is not automatically a linear intensity.
The positivity assumption means the usual one-sided diffuse clamp is inactive for the observations being used. A cast shadow or a light behind the surface can invalidate the linear equation; the model must be revised or those observations excluded. Unknown per-light gains cannot simply be ignored. We work with one scalar channel; color, interreflection and spatially varying illumination require additional modeling.
Solving for scaled normal
Let , let have rows , and write . The unknown has three components: two for the direction of the unit normal and one for .
Proposition 8.1.2
(Exact recovery from three lights)
Let with invertible, and let the data be exact, with and . Then , and . If , no unit normal is determined.
Proof
(Invert and normalize)
Multiplying by gives . Since is a unit vector and , and . When , normalization would divide by zero, and every unit satisfies with , so the equations do not select one.
With more than three observations the exact data need not be consistent, and one fits by least squares.
Theorem 8.1.3
(More observations and the rank condition)
For , the unweighted least-squares problem has a unique minimizer exactly when has rank three. In that case it satisfies
If , one may form and . These quantities are estimates, not proof that all model assumptions hold.
Proof
(Normal equations and the null space of the lights)
The objective is convex with gradient , so its minimizers are exactly the solutions of . For any vector , . If has rank three, then only for , so is positive definite and invertible, and the normal equations have exactly one solution. If the rank is smaller, some nonzero has ; then , so every minimizer gives another minimizer , and uniqueness fails.
Proof X-ray
Lighting must span three directions
The rank hypothesis enters at one point: it turns for nonzero into invertibility of . The data never affect uniqueness; they decide only where the minimizer lies. A null vector of is a direction in which every observation is blind, because every light is perpendicular to it, so adding to changes no predicted intensity. Full rank guarantees an algebraic solution. It does not guarantee that noisy data obey the positivity or reflectance assumptions; those are hypotheses of the model, not consequences of the theorem.
The rank condition has a direct geometric reading, because each observation is one linear equation in .
Concept lensGeometric
Observations as planes in the space of scaled normals
With , observation says that lies on the plane , whose normal direction is . Three planes with independent normals meet in exactly one point, which is . If the lights are coplanar or repeated, the plane normals are dependent, so for consistent data the planes meet in a line or a whole plane, and for inconsistent data they may have no common point. When they meet, the common points are the vectors with in the null space of .
A three-light calculation
Worked example 8.1.4
(A non-axis-aligned normal)
Choose
and exact observations
The row matrix has determinant , so its rows are independent. Solving gives , with and . The dot products with the three lights are , and , all positive. Multiplying by recovers the three observations. These checks matter: an invertible matrix alone would not establish the one-sided illumination assumption.
Worked example 8.1.5
(Three lights in one plane)
Take , and , and observations . All three lights have zero -component, so the second column of vanishes and has rank two even though there are three distinct lights. Both and reproduce exactly. The first gives , ; the second gives , , with positive dot products , and with , and . The data cannot say which is correct, because they are blind to the direction.
Worked example 8.1.6
(Least squares with a repeated light)
Take four lights , , and observations . Then and , so : the repeated light contributes the average of its two observations. The residuals are in the last two observations and zero in the first two. Then and , with positive dot products with the four lights. Repeating a light averages noise, but it does not add a new direction.
Counterexample mode
Repeated lighting cannot separate normal and reflectance
Suppose every light is and every observation is . The pair , fits, as does , . Both normals are unit and the relevant dot products are positive, yet the normals differ. Repeating the same lighting direction adds observations but not independent directional information.
Common mistake
(Two lights suffice because a normal has two degrees of freedom)
The unit normal has two degrees of freedom, but the unknown also contains the reflectance , so there are three unknowns. With , and , both and fit exactly, because is perpendicular to both lights. They give different normals, and , with and . Two images cannot identify a normal and a reflectance together.
Limits and uncertainty
Small measurement residuals do not guarantee a correct normal if the calibrated lights, correspondence or diffuse model are wrong. Even with full rank, nearly dependent rows can make the solution sensitive to noise; the smallest singular value of quantifies a weakly constrained direction. One can inspect residuals and conditioning and reject dark or shadowed samples, but doing so does not by itself prove an unbiased result. If a fitted is nearly zero, its normalized direction is unstable. Constraints on reflectance or one-sided illumination require an explicitly constrained method rather than assuming an unconstrained least-squares solution respects them.
A normal is local orientation, not a complete depth map. Integrating a field of estimated normals into a surface needs further consistency conditions, boundary choices and an integration method.
Exercises
Quick check
(For the three-light example, verify the first and third observed intensities from .)
Take dot products of with and .
Solution · Answer
and , as stated.
Quick check
(If three calibrated lighting rows are all , what is their rank? Can repeated observations alone determine all three components of ?)
Which components of appear in the equations?
Solution · Answer
The row rank is one. The observations constrain only the third component, so the first two are not determined by these equations.
Quick check
(If a full-rank exact solution is , can the unit normal be recovered by dividing by ?)
What is here?
Solution · Answer
No. Its norm is zero, and the normalization is undefined; the model does not determine the normal from that scaled vector.
Quick check
(Explain why an image taken under a cast shadow should not automatically be inserted into the unclamped linear equation.)
Compare the illumination the model assumes with what a shadowed point receives.
Solution · Answer
A shadowed point may receive less direct illumination than the model assumes. Its equation is then not necessarily with only the stated small-error interpretation; exclude it or extend the illumination model.
Further study
Unknown illumination, non-Lambertian reflectance, shadows, noise-aware estimation and integration of normal fields require additional analysis. This unit does not implement a photometric-stereo system or certify any particular reconstructed surface.