Metric Error Assessment Regarding Geometric 3D Reconstruction of Transparent Surfaces via SfM Enhanced by 2D and 3D Gaussian Splatting

TL;DR

This study compares methods for 3D reconstruction of transparent objects, finding that 2D Gaussian splatting provides the best balance of accuracy and detail.

Contribution

The paper introduces a practical methodology for assessing metric accuracy in 3D transparent object reconstruction using Gaussian splatting techniques.

Findings

2D Gaussian splatting captures fine surface and internal details with better geometric consistency than 3DGS and photogrammetry.

Photogrammetry fails to reconstruct transparent objects entirely, while 3DGS introduces surface artifacts affecting metric reliability.

Ground truth comparisons confirm that 2DGS offers the best trade-off between accuracy and appearance for transparent surfaces.

Abstract

This research investigates the metric accuracy of 3D transparent object reconstruction, a task where conventional photogrammetry often fails. The topic is especially relevant in cultural heritage (CH), where accurate digital documentation of glass and transparent artifacts is important. The work proposes a practical methodology using existing tools to verify metric accuracy standards. The study compares three methods, conventional photogrammetry, 3D Gaussian splatting (3DGS), and 2D Gaussian splatting (2DGS), to assess their ability to produce complete and metrically reliable 3D models suitable for measurement and geometric analysis. A transparent glass artifact serves as the case study. Results show that 2DGS captures fine surface and internal details with better geometric consistency than 3DGS and photogrammetry. Although 3DGS offers high visual quality, it introduces surface artifacts that affect metric reliability. Photogrammetry fails to reconstruct the object entirely. The study highlights that visual quality does not ensure geometric accuracy, which is critical for measurement applications. In this work, ground truth comparisons confirm that 2DGS offers the best trade-off between accuracy and appearance, despite higher computational demands. These findings suggest extending the experimentation to other sets of images featuring transparent objects, and possibly also reflective ones.

Figures