Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comparative Study of Neural Surface Reconstruction for Scientific Visualization

Siyuan Yao, Weixi Song, Chaoli Wang

TL;DR

This work provides a comprehensive benchmark comparing ten neural surface reconstruction methods across radiance-field and neural implicit approaches, with a focus on scientific visualization from multi-view renderings. It demonstrates that distance-function representations—particularly SDFs for closed surfaces and UDFs for open surfaces—can substantially improve surface accuracy and smoothness, highlighting NeuS2 for closed surfaces and NeUDF as a promising open-surface candidate despite extraction challenges. The study uses a nine-dataset benchmark with standardized training and evaluation, offering practical guidance on method selection and emphasizing the need for improved open-surface extraction methods. By releasing their dataset and insights, the authors aim to accelerate advancement in neural surface reconstruction for visualization and analysis in scientific contexts.

Abstract

This comparative study evaluates various neural surface reconstruction methods, particularly focusing on their implications for scientific visualization through reconstructing 3D surfaces via multi-view rendering images. We categorize ten methods into neural radiance fields and neural implicit surfaces, uncovering the benefits of leveraging distance functions (i.e., SDFs and UDFs) to enhance the accuracy and smoothness of the reconstructed surfaces. Our findings highlight the efficiency and quality of NeuS2 for reconstructing closed surfaces and identify NeUDF as a promising candidate for reconstructing open surfaces despite some limitations. By sharing our benchmark dataset, we invite researchers to test the performance of their methods, contributing to the advancement of surface reconstruction solutions for scientific visualization.

A Comparative Study of Neural Surface Reconstruction for Scientific Visualization

TL;DR

This work provides a comprehensive benchmark comparing ten neural surface reconstruction methods across radiance-field and neural implicit approaches, with a focus on scientific visualization from multi-view renderings. It demonstrates that distance-function representations—particularly SDFs for closed surfaces and UDFs for open surfaces—can substantially improve surface accuracy and smoothness, highlighting NeuS2 for closed surfaces and NeUDF as a promising open-surface candidate despite extraction challenges. The study uses a nine-dataset benchmark with standardized training and evaluation, offering practical guidance on method selection and emphasizing the need for improved open-surface extraction methods. By releasing their dataset and insights, the authors aim to accelerate advancement in neural surface reconstruction for visualization and analysis in scientific contexts.

Abstract

This comparative study evaluates various neural surface reconstruction methods, particularly focusing on their implications for scientific visualization through reconstructing 3D surfaces via multi-view rendering images. We categorize ten methods into neural radiance fields and neural implicit surfaces, uncovering the benefits of leveraging distance functions (i.e., SDFs and UDFs) to enhance the accuracy and smoothness of the reconstructed surfaces. Our findings highlight the efficiency and quality of NeuS2 for reconstructing closed surfaces and identify NeUDF as a promising candidate for reconstructing open surfaces despite some limitations. By sharing our benchmark dataset, we invite researchers to test the performance of their methods, contributing to the advancement of surface reconstruction solutions for scientific visualization.
Paper Structure (8 sections, 1 equation, 4 figures, 5 tables)

This paper contains 8 sections, 1 equation, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Radiance Fields and Implicit Surfaces
  4. Comparison and Discussion
  5. Datasets, Training, and Metrics
  6. Comparison Results
  7. Summary
  8. Concluding Remarks

Figures (4)

  • Figure 1: Inferred neural rendering images (upper-left) and rendering images of reconstructed surfaces (lower-right) of Nyx, Tangaroa, and vortex generated by (a) NeRF, (b) TensoRF, (c) Instant-NGP, and (d) NeuS. (e) shows the GT results.
  • Figure 2: Inferred neural rendering images (upper-left) and rendering images of reconstructed surfaces (lower-right) of aorta, combustion, and supernova generated by (a) IDR, (b) NeuS, (c) NeuS2, and (d) Neuralangelo. (e) shows the GT results.
  • Figure 3: Cut-through rendering of reconstructed supernova surfaces generated by (a) IDR, (b) NeuS, (c) NeuS2, and (d) Neuralangelo. (e) shows the GT results.
  • Figure 4: Inferred neural rendering images (upper-left) and rendering images of reconstructed surfaces (lower-right) of five critical points, Marschner-Lobb, and solar plume generated by (a) NeuS, (b) NeAT, (c) NeUDF, and (d) NeuralUDF. (e) shows the GT results.