CiGS: Conduction in Gaussian Splats for Interior Temperature Field from Thermal Images
Organizations: ETH Zürich, Switzerland · Carnegie Mellon University, United States · Google DeepMind, United States
Abstract
We present a method for inferring the interior temperature distribution of a solid object from registered multi-view visible-spectrum and thermal images, even when the thermal views cover only part of its surface. We formulate steady-state heat conduction as a Laplace boundary value problem and solve it with a direct boundary integral equation discretized over a point cloud: a Gaussian reconstruction of the visible-spectrum images supplies the boundary geometry, and the thermal images supply the boundary data. We do not build surface or volume meshes at any stage in our method. Instead, we sample collocation points from rendered depth without the pixels at depth discontinuities, and use a density subtraction that keeps the interior evaluation accurate up to the surface. The whole chain is differentiable, and we use it to recover the pose and depth of heating pipes, insulation defects, and the shape of cracks from thermal images. We validate our method through experiments on boundary value problems with Dirichlet-only and mixed Dirichlet and Robin boundaries, using synthetic and real captured data on scenes with challenging topology. These experiments show that our method matches analytical solutions and has better accuracy than FEM solvers.
Figures & tables
| full coverage | single view | |||||
| interior | surface | interior | ||||
| scene | ours | mesh | ours | mesh | ours | mesh |
| Armadillo | 0.00073 | 0.00083 | 0.00647 | 0.01057 | 0.00450 | 0.00622 |
| Botijo | 0.00105 | 0.00172 | 0.00997 | 0.01493 | 0.00453 | 0.00923 |
| Bunny | 0.00062 | 0.00082 | 0.00516 | 0.00758 | 0.00208 | 0.00377 |
| Hand | 0.00083 | 0.00095 | 0.00465 | 0.00971 | 0.00173 | 0.00471 |
| full coverage | single view | ||
| variant | interior | interior | surface |
| full pipeline (ours) | 0.00167 | 0.00581 | 0.00820 |
| without the depth-slope filter | 0.00309 | 0.00709 | 0.00988 |
| maps from 2DGS instead of RaDe-GS | 0.00290 | 0.00783 | 0.01267 |
| without the membership filter | 0.00199 | 0.00610 | 0.00850 |
| without density subtraction | 0.03473 | 0.04597 | 0.00820 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Inverse Problems
Inverse problems infer hidden causes, parameters, or signals from indirect, noisy measurements using physical models, priors, and statistical methods.
See all 266 papers