The paper presents a theory and parallel algorithm for estimating surface shape from a single image using only local shading information, without any a‑priori knowledge of illumination, reflectance, or scene geometry. It proves that the image‑plane component of surface orientation (tilt) can be obtained exactly from the second derivatives of image intensity, and that a maximum‑likelihood estimate of the remaining depth component (slant) can be derived under a uniform curvature prior. The author formalizes these results in two propositions (Tilt of the Surface and Estimation of Slant), derives the necessary mathematical relationships, and implements a local, parallel “shape algorithm” on an MIT LISP machine. The algorithm is evaluated on synthetic images of known geometry (sphere, ellipsoid) where it recovers shape to within 0.01 % error, and on natural images (log, rock, electron‑microscope samples) where the generated relief maps closely match the actual object shapes. The work contrasts with earlier shape‑from‑shading methods that require global constraints or a‑priori scene knowledge, demonstrating that local shading cues alone are sufficient for high‑quality shape inference.
Topics
Exact Determination of Surface Tilt from Second‑Derivative Shading
Derivation and proof that the direction of maximal second‑derivative intensity (the spread term) directly yields the image‑plane component of the surface normal (tilt) for smooth second‑order surfaces.
Maximum‑Likelihood Estimation of Surface Slant under Uniform Curvature Prior
Formulation of a normalized Laplacian of intensity relationship, leading to a closed‑form ML estimator for the z‑component of the surface normal (slant) assuming a uniform distribution of surface curvatures.
Parallel Local Shape‑From‑Shading Algorithm (Shape Algorithm)
Implementation details of the algorithm that computes tilt and slant locally at each pixel using convolution masks for second‑derivative and Laplacian calculations, without global integration or scene priors.
Synthetic Image Evaluation: Near‑Perfect Recovery of Known Geometry
Experiments on rendered sphere and ellipsoid images showing that the algorithm recovers shape to within 0.01 % error, demonstrating the accuracy of the local approach under ideal conditions.
Natural Image Evaluation: Relief Maps for Real‑World Objects
Application of the algorithm to digitized photographs of a log, a rock, and electron‑microscope images, with generated relief maps that closely match perceived object shapes, illustrating robustness to noise and complex illumination.
Contrast with Prior Shape‑From‑Shading Methods
Discussion of earlier work (Horn, Ikeuchi) that relies on global integration, known illumination, and reflectance, highlighting the novelty of a purely local, a‑priori‑free approach.
Assumptions, Error Analysis, and Limitations
Examination of the second‑order surface assumption, error bounds (δ³), handling of twisting surfaces, and the inherent undetermined degree of freedom that remains despite the tilt and slant propositions.