Objects as Volumes: A Stochastic Geometry View of Opaque Solids

Bailey Miller, Hanyu Chen, Alice Lai, Ioannis GkioulekasCVPR 2024 · DOI · arXiv

Abstract

This paper establishes a rigorous mathematical foundation for representing opaque solids as volumes using stochastic geometry. By proving that volumetric transport is valid under Markovian assumptions, the authors derive a reciprocal attenuation model that improves 3D reconstruction quality and physical consistency compared to existing heuristic methods like NeRF, NeuS, and VolSDF.

Introduction

Neural rendering techniques, most notably NeRF, have revolutionized 3D reconstruction by treating scenes as continuous volumetric densities. However, a conceptual mismatch exists: volumetric rendering was originally designed for participating media like smoke or fog, whereas opaque solids possess sharp surfaces where light stops abruptly. Previous methods relied on heuristic "hacks"—such as using Sigmoids or ReLUs to map distance functions to density—which often violate physical laws like reciprocity and reversibility. This work addresses this gap by asking: What is the mathematical underpinning for modeling an opaque object as a volume, and how can we do it correctly?

A Stochastic Geometry Framework for Volumes

The authors turn to stochastic geometry, representing an opaque solid as a random indicator function that is either "solid" or "empty" at any point. A central contribution is the Theorem of Exponential Transport, which proves that volumetric transport (exponential decay of light) is physically valid only if the object's geometry follows a continuous-time Markov process along a ray.

Mathematical pipeline for deriving volumetric properties from implicit distributions.

To bridge the gap between fuzzy volumes and sharp surfaces, the authors decompose the volume into an isotropic "density" and an anisotropic "projected area." This allows the model to behave like a flat surface near boundaries while remaining a uniform medium deep inside. They also provide a "corrected" integration for Signed Distance Functions (SDFs), fixing mathematical oversights in previous models like VolSDF.

The Importance of Reciprocity

A key insight of this work is the necessity of reciprocity—the principle that light should behave identically whether traveling from point A to B or vice versa. Previous state-of-the-art models like NeuS used ReLU-based terms that made surfaces invisible when viewed from the "inside out," breaking the symmetry required for consistent light transport.

Comparison of non-reciprocal (NeuS) vs. reciprocal (ours) transmittance models showing improved consistency.

By enforcing reciprocity through a new attenuation coefficient model, the authors demonstrate that physical plausibility leads directly to better geometric reconstruction. This consistency is essential for advanced global illumination effects where rays must be traced in multiple directions.

Experimental Results

The proposed method was evaluated against standard benchmarks including DTU, BlendedMVS, and NeRF Realistic Synthetic datasets. The "corrected" volumetric representation consistently outperformed existing methods:

  • Quantitative Accuracy: On the DTU dataset, the method achieved a mean Chamfer distance of 1.57, significantly lower than VolSDF (1.84) and NeuS (2.17).
  • Qualitative Fidelity: Reconstructions exhibited much cleaner surfaces and more accurate normals, particularly in challenging concave regions or areas with complex self-occlusions.

Conclusion and Future Work

While this theory provides a robust foundation for opaque solids, it currently excludes semi-transparent or refractive materials. Future research may explore non-exponential transport models based on different stochastic geometries, such as Gaussian processes, and the integration of this rigorous geometry with global illumination models to account for inter-reflections.

Summary updated

BibTeX
@inproceedings{miller2024objects,
  author = {Bailey Miller and Hanyu Chen and Alice Lai and Ioannis Gkioulekas},
  title = {{Objects as Volumes: A Stochastic Geometry View of Opaque Solids}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2024},
  doi = {10.1109/cvpr52733.2024.00017},
}
Figure