TL;DR
CiGS (Conduction in Gaussian Splats) is a mesh-free framework that predicts the internal 3D temperature of an object using only standard RGB and thermal photographs. By solving heat conduction physics directly on point clouds extracted from 3D Gaussian Splats, it bypasses the fragile and time-consuming process of 3D meshing while outperforming traditional Finite Element Methods (FEM) in both speed and accuracy.
What Problem Does This Paper Address?
Thermal cameras are excellent for seeing the "skin" of an object, but in many fields—industrial inspection, building energy audits, and electronics cooling—the most critical information is hidden inside. We want to know if a floor-heating pipe is leaking, if insulation has settled behind a wall, or if a structural component has an internal crack.
Traditionally, inferring interior temperatures requires two things: a precise 3D model (a mesh) and a physics solver (like FEM). However, generating a "simulation-ready" mesh from photos is notoriously difficult. Automated tools often produce meshes with holes, overlapping faces, or "thinned" structures that cause physics solvers to crash or yield wildly inaccurate results. The authors of CiGS ask: Can we skip the mesh entirely and solve the physics directly on the raw output of modern neural reconstruction?
Methods
CiGS bridges the gap between computer vision and thermal physics using a three-step pipeline:
- Geometry Reconstruction: The system uses RaDe-GS (a variant of 3D Gaussian Splatting) to reconstruct the object from visible-spectrum images. Instead of a mesh, it renders depth and normal maps to extract a dense, clean point cloud of "collocation points."
- Physics Formulation: Heat flow in a steady state follows the Laplace equation. The authors solve this using a Direct Boundary Integral Equation (BIE). Unlike FEM, which discretizes the entire volume, BIE only requires data on the surface (boundary), making it a perfect fit for point clouds.
- Numerical Innovations:
- Depth-Slope Filtering: A novel filter that identifies and removes "flying pixels" at depth discontinuities (like the edges of a thin strut), which would otherwise introduce massive errors into the physics simulation.
- Density Subtraction: A mathematical technique that handles the "singularity" problem—where temperature calculations usually blow up as you get closer to the surface. This allows for accurate temperature prediction right up to the object's skin.
- Differentiability: The entire chain is end-to-end differentiable. This allows the model to solve inverse problems: you can start with a surface thermal image and "backpropagate" to find the hidden location of a pipe or the shape of an internal crack.
Datasets
The authors evaluated CiGS on a diverse range of data:
- Synthetic Dataset: 18 complex meshes categorized into solids (e.g., Armadillo, Bunny), thin structures (e.g., Voronoi lattices, spiral ribbons), and industrial parts.
- Real-World Capture: A custom-built rig featuring a plaster-cast teapot on a hot plate. The teapot was embedded with three internal temperature probes to provide ground-truth data for validation.
- Benchmarks: The method was compared against a state-of-the-art mesh-based pipeline using FEniCSx (a popular FEM solver) and various image-to-mesh reconstruction tools like MILo and TopoSurfel.
Key Insights and Ideas
- Meshing is the Bottleneck: The paper demonstrates that for complex topologies (like thin struts), the "meshing tax" is high. Traditional FEM solvers failed to provide valid solutions for up to 44% of the interior volume in complex shapes because the reconstructed meshes were too poor. CiGS, being mesh-free, remained robust.
- Points > Elements: By treating the surface as a collection of independent points rather than connected triangles, the system handles complex topology (holes, thin shells) much more gracefully.
- Physics-Aware Reconstruction: Because the solver is differentiable, it doesn't just "guess" what's inside; it optimizes the internal parameters (like the depth of a pipe) until the simulated surface temperature matches the actual thermal images.
Key Results
- Accuracy: On synthetic scenes, CiGS achieved a mean absolute error of just 0.17% of the temperature range, significantly lower than mesh-based baselines.
- Speed: For the "Armadillo" scene, the CiGS physics solve took only 31 seconds, compared to 210 seconds for the mesh-based pipeline (where 134 seconds were wasted just on tetrahedralization).
- Real-World Validation: In the teapot experiment, CiGS predicted internal temperatures that agreed with physical probe readings to within 1.4°C on average.
- Inverse Problem Success: The model successfully recovered the 3D pose of a hidden heating pipe and the shape of a sub-surface crack in a propeller hub that was completely invisible in RGB photos.
Limitations and Future Work
- Steady-State Assumption: The current model assumes the object is in a thermal steady state (temperatures aren't changing over time). Real-world scenarios often involve transient heat flow.
- Homogeneous Materials: The core Laplace formulation assumes the object is made of a single, uniform material. While the authors demonstrated a workaround for layered walls, a general multi-material solver is a goal for future work.
- Surface Radiance: The model assumes a purely emissive surface. Incorporating reflections (non-zero thermal reflectance) would improve accuracy for metallic or glossy objects.
- Reconstruction Quality: The accuracy of the physics solve is still tethered to the quality of the Gaussian Splat reconstruction. Improvements in 3DGS for thin structures will directly translate to better thermal predictions.
Summary updated