Shape from Darkness: Deriving Surface Information from Dynamic Shadows

John R. Kender, Earl SmithAAAI 1986

Abstract

Traditional computer vision techniques for 3D reconstruction often struggle with rough, non-smooth surfaces where specularities and complex textures dominate. This paper introduces "Shape from Darkness," a method that leverages the movement of shadows under a shifting light source to reconstruct surface geometry. By treating shadows as primary data rather than noise, the approach achieves high accuracy on fractal-like surfaces without requiring prior knowledge of material reflectance.

Introduction

Most "shape from X" methodologies, such as shape from shading or photometric stereo, rely on smooth luminance gradients. These methods frequently fail on "real-world" textures like crumpled foil or jagged terrain, where sharp transitions between light and shadow are the defining features. The authors propose using these transitions—specifically self-shadowing and cast shadows—to extract geometric information under extreme lighting conditions.

Methodology

The core of the reconstruction process is the Suntrace, a representation that maps the specific light source angle (or "time") at which each surface point first transitions from shadow to illumination.

Point Classification and Geometric Constraints

The algorithm classifies surface points into three categories:

  • Minus points: Self-shadowed points where the local surface slope blocks the light.
  • Zero points: Points that are never shadowed, typically representing global peaks.
  • Plus points: Points within a cast shadow, obscured by a distant feature.

A critical insight is the "grazing" condition: at the moment of transition, the light ray is tangent to the surface (for self-shadows) or connects two peaks (for cast shadows). This provides a direct constraint on the surface derivative.

Figure 6: The grazing light condition establishes a direct link between the light source angle (t) and the surface slope at the transition point.

Discrete Reconstruction via Relaxation

For digital images, the problem is modeled as a constraint satisfaction task. Each pixel is bounded by eight equations defining its possible height range. A relaxation algorithm iteratively tightens these bounds, allowing height information to propagate across the surface until a consistent estimate is reached.

Experimental Validation

The method was tested on several 1D simulated datasets to evaluate its robustness and accuracy.

Datasets

  • Fractal Mountain: A self-similar curve 300 points wide, designed to test performance on highly textured, non-smooth geometry.
  • Smooth Semicircle: Used to verify that the method remains compatible with traditional smooth surfaces.
  • Sparse Suntrace: A test using only four light source positions to assess robustness under limited data.
Figure 3: The self-similar fractal 'mountain' used as a primary test case for reconstructing complex, non-smooth geometry.

Key Results

The algorithm demonstrated high precision, achieving a cumulative error of less than 0.2 pixels per point on the fractal mountain. Even with sparse data (45-degree jumps in light angle), the reconstruction remained mathematically consistent and recognizable. On smooth surfaces, the error was kept below 0.5 units, proving the method's versatility.

Discussion and Conclusion

"Shape from Darkness" offers a robust alternative to shading-based methods by utilizing binary shadow signals, which are less susceptible to camera noise and independent of complex reflectance models. While currently demonstrated in 1D, the principles of information propagation and grazing constraints provide a foundation for 2D surface reconstruction. Future work aims to integrate these techniques into robotic environments where known boundary conditions can further refine surface estimates.

Summary updated

BibTeX
@inproceedings{kender1986shape,
  author = {John R. Kender and Earl Smith},
  title = {{Shape from Darkness: Deriving Surface Information from Dynamic Shadows}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure