A Geometric Perspective on Structured Light Coding

Mohit Gupta, Nikhil NakhateECCV 2018 · DOI

Abstract

Structured light (SL) 3D scanning often struggles in low signal-to-noise ratio (SNR) conditions, such as dark objects or high ambient light. This paper introduces a novel geometric framework that treats SL coding schemes as curves in a high-dimensional measurement space. By maximizing the length of these curves while avoiding self-intersections, the authors developed Hamiltonian SL coding. This approach achieves up to an order of magnitude higher precision than standard sinusoid methods using as few as three images, even in the presence of complex interreflections.

The Challenge of Precision in Structured Light

Structured Light (SL) 3D imaging works by projecting known patterns onto a scene and capturing how they deform. By analyzing these deformations, we can triangulate the 3D shape of objects. However, current SL methods face a "Goldilocks" problem: laser scanning is precise but slow; single-shot methods are fast but lack detail; and multi-pattern methods like sinusoid phase-shifting fail when objects are dark (low albedo), the room is bright (high ambient light), or light bounces between surfaces (interreflections).

Structured light setup (a) and the resulting projected patterns and intensity profiles (b).

A Geometric Framework for Coding

The core innovation of this work is a geometric framework that views a sequence of $K$ projected patterns as a curve in a $K$-dimensional "measurement space."

The Coding Curve and Precision Metric

If you project 3 patterns, every point on the projector can be represented as a 3D coordinate $(I_1, I_2, I_3)$. As you move across the projector's columns, these coordinates trace a path. The authors prove that the precision of a 3D scan is directly proportional to the length of this coding curve ($\Lambda$). A longer curve spreads out the points in measurement space, making them easier to distinguish even when noise "blurs" the data.

Visualization of the coding curve in the 3D measurement space (d), where the noise ellipsoid illustrates how uncertainty affects point localization.

Hamiltonian Coding

To maximize this length without the curve crossing itself (which would cause ambiguity), the authors look to graph theory. They use Hamiltonian cycles on hypercube graphs—paths that visit every corner of a cube exactly once. These paths create "trapezoidal" patterns that are much longer than the "circular" paths created by standard sinusoid patterns.

To handle light bouncing (interreflections), they propose two variants:

  • Antipodal Hamiltonian Coding: Based on Gray codes, ensuring patterns have high spatial frequencies.
  • Micro Hamiltonian Coding: A multi-frequency approach similar to "micro phase-shifting" but using Hamiltonian building blocks.

Experimental Evaluation

The authors evaluated their method using a physical laboratory setup consisting of a Canon T5i DSLR camera and an Epson 3LCD projector. Quantitative tests on a diffuse white planar surface measured depth error across varying signal-to-noise levels. Qualitative tests involved complex objects like a Buddha statue, black lava rock (extremely low albedo), a white glossy bowl (high interreflections), and a candle (subsurface scattering).

Key Results and Insights

  • 10x Precision Gain: In low SNR scenarios, Hamiltonian coding achieved up to 10x higher precision than sinusoid phase-shifting.
  • Efficiency: High-quality 3D shapes can be recovered with as few as three images.
  • Robustness: The method successfully reconstructed fine details on black lava rock, where standard methods failed. It also reduced "bleeding" errors in concave objects caused by interreflections.
  • Distance Preservation: A key requirement for a good code is being "distance-preserving." While some curves (like Hilbert curves) are long, they fail because they bring distant points too close together, leading to massive errors when noise is present.

Limitations and Future Work

While Hamiltonian cycles are significantly better than existing patterns, they are not yet mathematically proven to be the absolute optimal patterns. Furthermore, while robust, these patterns alone cannot solve extreme cases like mirrors without additional optical techniques. Future work may involve integrating scene priors and imaging system characteristics directly into the code design process to push precision even further.

Summary updated

BibTeX
@inproceedings{gupta2018geometric,
  author = {Mohit Gupta and Nikhil Nakhate},
  title = {{A Geometric Perspective on Structured Light Coding}},
  booktitle = {European Conference on Computer Vision},
  year = {2018},
  doi = {10.1007/978-3-030-01270-0\_6},
}
Figure