A Theory of Refractive Photo-Light-Path Triangulation

Visesh Chari, Peter F. SturmCVPR 2013 · DOI

Abstract

Traditional 3D reconstruction of transparent objects often requires complex setups and multiple measurements to resolve geometric ambiguities. This paper presents a novel approach that integrates geometric light-path triangulation with radiometric cues derived from the Fresnel equations. By analyzing the intensity ratio of reflected and refracted light, the proposed method significantly reduces the number of required measurements, enabling high-fidelity reconstruction from a single static setup.

Introduction: The Challenge of Transparency

Reconstructing the 3D geometry of transparent objects, such as glassware or plastic containers, remains a significant challenge in computer vision. Unlike opaque surfaces, transparent materials lack a consistent texture; their appearance is dominated by the reflection and refraction of the surrounding environment.

Previous state-of-the-art methods relied on "Light-Path Triangulation," which treats light as a series of linear segments. However, this purely geometric approach is data-intensive, typically requiring at least two distinct viewpoints or light source positions to triangulate a single surface point. This paper addresses whether the "radiometric signature"—the specific intensity changes occurring at the interface—can simplify this process.

The geometric setup for light-path triangulation, showing the path from a light source through the object to the camera.

Methodology: Fusing Geometry and Radiometry

The core innovation lies in the simultaneous application of Snell’s Law for geometry and the Fresnel Equations for radiometry.

The Radiometric Cue

When unpolarized light strikes a transparent surface, it splits into reflected ($I_r$) and refracted ($I_t$) components. The ratio of reflected intensity to incident intensity is a deterministic function of the incident angle ($\theta_1$) and the material's refractive index. By treating the intensity ratio as a known variable, the system of equations for surface reconstruction becomes solvable with only a single reference point on a light source.

Physics of unpolarized light interacting with a refractive surface, producing partially polarized reflected and refracted rays.

Mathematical Formulation

The authors demonstrate that this intensity relationship can be modeled as a quartic polynomial in $\sin^2 \theta_1$. This "Single-Point Triangulation" eliminates the need for moving hardware components, as a single image from a static setup can theoretically suffice if the light source's radiance is known.

Experimental Setup and Results

The researchers validated their theory using a calibrated DSLR camera and a CRT monitor (chosen for its unpolarized light emission). Gray-coded patterns were used to establish correspondences between the monitor pixels and camera pixels.

Key Findings

  • Accuracy: In controlled experiments involving water surfaces, the method achieved a mean squared error of approximately 0.1 cm.
  • Detail: Reconstruction of complex objects, such as a plastic Fanta bottle, captured fine-scale surface details and scratches that are often lost in purely geometric or smoothed models.
  • Robustness: Synthetic tests showed that incorporating radiometric data provides more accurate surface normals than geometric triangulation alone, especially when light sources are positioned close to the object.

Conclusion and Future Work

By leveraging the physics of light intensity, this research transforms what was previously considered "noise" (specular reflections) into a powerful signal for 3D reconstruction. While the current model assumes a single "bounce" and requires prior knowledge of the refractive index, it paves the way for more efficient acquisition systems. Future work will explore multi-bounce scenarios and the complexities introduced by polarized light sources like modern LCDs.

Summary updated

BibTeX
@inproceedings{chari2013theory,
  author = {Visesh Chari and Peter F. Sturm},
  title = {{A Theory of Refractive Photo-Light-Path Triangulation}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2013},
  doi = {10.1109/cvpr.2013.189},
}

Loading related content...

Figure