TSPE-GS: Probabilistic Depth Extraction for Semi-Transparent Surface Reconstruction via 3D Gaussian Splatting

Zhiyuan Xu, Nan Min, Yuhang Guo, Tong WeiarXiv 2025 · DOI · arXiv

Abstract

3D Gaussian Splatting (3DGS) has significantly advanced real-time 3D reconstruction, yet it remains limited when handling semi-transparent surfaces. Standard 3DGS methods typically estimate a single depth value per pixel, leading to "depth confusion" and artifacts in scenes with glass or plastic. We present TSPE-GS (Transparent Surface Probabilistic Extraction), a framework that treats depth as a multi-modal distribution. By identifying multiple probability peaks along a viewing ray, TSPE-GS can accurately reconstruct both transparent outer shells and the objects hidden within them. Our method acts as a plug-and-play module for existing 3DGS backbones, demonstrating superior performance on transparent reconstruction benchmarks.

Introduction

The core challenge in reconstructing semi-transparent scenes lies in the multi-layered nature of the geometry. When a ray passes through a transparent surface to hit an internal object, the captured pixel contains information from at least two distinct depths. Standard 3DGS pipelines often compute an "expected depth"—a weighted average—which often falls in empty space between the two actual surfaces. This results in "ghostly" reconstructions or the complete omission of internal geometry.

Methodology

TSPE-GS addresses this by shifting from single-point depth estimation to a probabilistic approach consisting of three main stages:

1. Probabilistic Depth Modeling

Instead of a single depth value, TSPE-GS samples transmittance values along each viewing ray to construct a Probability Density Function (PDF) of depth. This PDF captures the multi-modal nature of the scene, where peaks represent likely surface locations.

2. Peak Location Calculation

To efficiently identify surface locations without the prohibitive cost of per-pixel peak detection, the authors utilize Kernel Density Estimation (KDE). They demonstrate that mean depth values at specific transmittance thresholds can effectively pinpoint these physical surface peaks.

3. Progressive TSDF Fusion

Fusing multiple depth layers into a Truncated Signed Distance Function (TSDF) can cause interference. TSPE-GS employs a progressive strategy: the outermost surfaces are reconstructed and "frozen" first. Subsequent inner layers are then reconstructed, ensuring they do not corrupt the established high-quality exterior shell.

Experiments and Results

The method was evaluated on several datasets, including $\alpha$Surf (translucent objects) and a new "Bottleship" dataset featuring occluded objects.

Comparison of TSPE-GS integrated with 2DGS and RaDeGS backbones against ground truth (GT) for a transparent scene.

As shown in the figure above, standard methods like 2DGS and RaDeGS fail to capture the internal monkey figurine, either distorting the outer container or ignoring the interior entirely. The addition of the TSPE module allows these backbones to recover the internal geometry with high fidelity.

Quantitatively, TSPE-GS achieved a Chamfer Distance (CD) of 0.82 on $\alpha$Surf, outperforming the previous state-of-the-art (GOF at 1.13). On the Bottleship dataset, it reached a CD of 0.88, a significant improvement over PGSR’s 1.58.

Discussion and Conclusion

TSPE-GS demonstrates that the geometric information for transparent surfaces is already present within 3DGS models; the key is how that information is extracted. By moving to a probabilistic depth framework, we can resolve the ambiguity of multi-layered scenes. While currently a post-processing extraction method, future work could integrate this probabilistic awareness directly into the 3DGS training loop to better handle complex optical phenomena like refraction.

Summary updated

BibTeX
@misc{xu2025tspe,
  author = {Zhiyuan Xu and Nan Min and Yuhang Guo and Tong Wei},
  title = {{TSPE-GS: Probabilistic Depth Extraction for Semi-Transparent Surface Reconstruction via 3D Gaussian Splatting}},
  year = {2025},
  eprint = {2511.09944},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.09944},
}
Figure