Computing the Quality of Isosurface-Based Visualizations Using a Feed-Forward Model for Single-Image Reconstruction

Authors

Li-Ta Lo (Los Alamos National Lab), Christian Heine (Leipzig University), Roxana Bujack (Los Alamos National Laboratory)

Presentation

Session

Big Data, Bigger Physics

Time

Friday, Nov 13, 08:36 – 08:45 (US/Eastern) · session 08:00 – 09:30

Location

Hall America south

Keywords

Scientific visualization, machine learning, isosurface.

Abstract

Automatic evaluation of visualization quality enables optimization of visualization parameters without human intervention. In isosurface visualization, prior work estimates quality by reconstructing a 3D surface from multiple images using NeRF, followed by volumetric reconstruction and comparison to the original data. In this paper, we replace the NeRF-based multi-view reconstruction with a diffusion-based single-image 3D reconstruction model. This formulation removes the need for auxiliary views and aligns the evaluation setting with single-image visualization scenarios. Because the reconstructed geometry is not produced in the original coordinate system, we introduce an explicit alignment step to resolve global pose and scale ambiguities. Experimental results on benchmark datasets show that the proposed approach reduces surface reconstruction error and improves runtime relative to the NeRF-based pipeline. These results suggest that strong learned priors provide a more reliable reconstruction backend for reconstruction-based evaluation while enabling more efficient workflows.

For Practitioners

Simulation scientists, scientific visualization practitioners, developers of autonomous scientific visualization systems.