The paper addresses the fundamental limitation of traditional shape‑from‑shading and shape‑from‑texture techniques, which assume smooth surfaces and therefore fail on the rough, crumpled surfaces common in natural scenes. Pentland proposes a fractal surface model—specifically spatially isotropic fractal Brownian surfaces—as a unified representation that captures both smooth (low fractal dimension) and rough (high fractal dimension) surfaces. He proves that the fractal dimension of a surface’s normal vector equals the fractal dimension of the image intensity under Lambertian reflectance, enabling direct estimation of surface roughness from image data. Human perception experiments show a near‑perfect correlation (0.98) between perceived roughness and fractal dimension, validating the model’s perceptual relevance. From this model he derives a robust texture measure, the expected magnitude of the second derivative of intensity normalized by intensity, E(|d²I/I|), which is largely illumination‑invariant and can be used to infer surface orientation (slant and tilt) via regional estimation, handling both 2‑D texture and 3‑D texture. An algorithm is presented that computes regional minima (α) and maxima (β) of this measure, derives the surface normal’s z‑component, and integrates slant and tilt to produce a relief map. The method gracefully degrades to classic shape‑from‑shading when texture diminishes, as demonstrated on a real image of Tuckerman’s Ravine, where the estimated relief matches topographic data with correct slopes up to 60°. The work unifies shading and texture under a single fractal framework, offering a principled way to extend shape estimation to natural, rough surfaces.
Topics
Motivation: Limitations of Smooth‑Surface Shape Estimation
Discusses how existing shape‑from‑shading and shape‑from‑texture methods rely on smoothness assumptions, which break down for natural, rough surfaces.
Fractal Surface Model: Spatially Isotropic Fractal Brownian Surfaces
Introduces the fractal Brownian function, defines fractal dimension D = 2‑H, and presents the spatially isotropic fractal surface model as a realistic description of natural terrain.
Proposition 1: Image Intensity Fractality Mirrors Surface Normal Fractality
States and justifies that under Lambertian reflectance, the fractal dimension of the image intensity surface equals that of the surface normal components.
Human Perception Validation of Fractal Roughness Measure
Presents experimental results showing a 0.98 correlation between subjects' roughness ratings and the fractal dimension of 1‑D curves and 2‑D surfaces.
Derivation of the Robust Texture Measure E(|d²I/I|)
Derives a second‑derivative based texture statistic that is illumination‑invariant and proportional to the average magnitude of the surface normal’s second derivative.
Regional Estimation and Anisotropy Handling
Explains how regional (median) estimates of the texture measure reduce error from local texture variation and provide a test for anisotropy via differing fractal dimensions in x and y.
Unified Shape‑from‑Texture‑and‑Shading Algorithm
Details the step‑by‑step algorithm: compute α = min E(|d²I/I|), β = E(|∇²I/I|), derive z_N, obtain slant and tilt, and integrate to a relief map.
Application to Tuckerman’s Ravine and Performance Evaluation
Shows the algorithm applied to a real photographed ravine, comparing the estimated relief to topographic data and noting accurate slope estimation up to 60°.