Acquiring Axially-Symmetric Transparent Objects Using Single-View Transmission Imaging

Jaewon Kim, Ilya Reshetouski, Abhijeet GhoshCVPR 2017 · DOI

Introduction

3D scanning transparent objects remains a significant challenge in computer vision. Traditional techniques like structured light or LiDAR are often ineffective because light refracts through transparent surfaces rather than reflecting off them. Existing solutions typically require complex multiview setups, specialized sensors, or invasive methods like immersing objects in index-matching fluids.

Researchers from Imperial College London have introduced a practical, single-view method to reconstruct the 3D geometry and refractive index of axially-symmetric transparent objects, such as wine glasses and carafes. By leveraging axial symmetry as a geometric prior, the system simplifies a high-dimensional reconstruction problem into a manageable optimization task using a commodity setup consisting of a single camera and an LCD monitor.

A variety of axially-symmetric transparent objects, including rotationally and n-fold symmetric glasses, evaluated in the study.

Methodology

The acquisition setup involves placing the target object between a camera and an LCD monitor. The monitor displays a sequence of 13 coded patterns, including linear gradients and high-frequency Gray codes, which the camera captures as they are distorted by the object.

The "Flip Image" and Segmentation

A core innovation of this work is the "flip image." By analyzing how light rays cross the object's vertical axis of symmetry, the system can automatically distinguish between solid and hollow sections. In solid glass sections, rays typically "flip" across the axis due to high refraction. In hollow sections, the additional air-glass interfaces alter this refraction profile. This physical insight allows for automatic segmentation of internal structures without manual labeling.

The reconstruction pipeline: (a) captured patterns, (b) deformation vector, (c) deformation amplitude, and (d) the 'flip image' used for segmentation.

Analysis-by-Synthesis Optimization

The reconstruction follows a sequential "analysis-by-synthesis" approach:

  1. Refractive Index Estimation: The system first determines the material's refractive index ($n$) by analyzing solid sections. It searches for the value of $n$ that minimizes the difference between observed ray deflections and those simulated via ray-tracing.
  2. Geometry Optimization: Once $n$ is fixed, the system optimizes the radius of the inner surfaces in hollow sections to match the captured transmission patterns.
  3. N-fold Symmetry: The method extends to "n-fold" symmetric objects (e.g., hexagonal tumblers) by modeling them with a polygonal base and optimizing face curvature using quadratic functions.

Results and Evaluation

The method was evaluated on a diverse set of real-world objects, including rotationally symmetric flutes, n-fold symmetric tumblers, and liquid-filled shot glasses.

  • Accuracy: Comparisons with high-precision physical measurements showed a mean error of only 0.35mm for rotationally symmetric objects and 0.86mm for n-fold symmetric objects.
  • Material Properties: The system successfully estimated the refractive indices of various materials, accurately identifying glass (~1.57) and liquids like vodka (~1.36).
  • Visual Fidelity: Models reconstructed with this method and rendered using path-tracing engines are visually indistinguishable from real photographs, accurately reproducing complex caustic patterns.

Conclusion and Future Work

This research demonstrates that axial symmetry is a powerful constraint for reconstructing transparent geometry from a single view. While highly effective for a wide class of common objects, the method is currently limited to perfectly symmetric shapes and does not account for surface reflections or hand-blown imperfections. Future improvements could involve reducing the optimization time—currently up to 90 minutes—and incorporating specular highlights to handle a broader range of material properties.

Summary updated

BibTeX
@inproceedings{kim2017acquiring,
  author = {Jaewon Kim and Ilya Reshetouski and Abhijeet Ghosh},
  title = {{Acquiring Axially-Symmetric Transparent Objects Using Single-View Transmission Imaging}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2017},
  doi = {10.1109/cvpr.2017.162},
}
Figure