Thermoxels

A voxel-based method to generate simulation-ready 3D thermal models

Etienne Chassaing
cetiennec@yahoo.com

EPFL

Florent Forest
florent.forest@epfl.ch

EPFL

Olga Fink
olga.fink@epfl.ch

EPFL

Malcolm Mielle
malcolm.mielle@schindler.com

Schindler EPFL Lab

Building (energy consumption) statistics

  • 85 to 95% still standing by 2050
  • Weighted annual energy renovation rate: 1%
  • Deep renovation: 0.2%

Need tools for targeting and monitoring of renovations.

Finite Element Analysis (FEA) for heat forward problem simulation.

  • Tedious and error-prone modeling

Aim

Simplify data collection and model generation

Research Question

Is it possible to use computer vision to simplify the generation of volumetric 3D models for Finite-Element Analysis (FEA)?

Existing 3D RGB+Thermal reconstruction

Take a sparse set of images of a building and estimate RGB and Thermal

  • NeRF for building facade novel view synthesis1 - No 3D model
  • Thermal Gaussian Splatting2 - Poorly defined interfaces between elements
  • Plenoxel3 - No thermal modality

Not possible to simulate using FEA

Thermoxels

First method to generate volumetric meshes compatible with FEA from a sparse set of RGB+Thermal images

Pose extraction

# Find poses on RGB with COLMAP
poses = COLMAP(images_rgb)

Voxel optimization

poses = COLMAP(images_rgb)
# optimize color and temperature on voxels
thermoxels = Voxels(scene_size)
while i <= max_iteration:
 rgb, temperature = thermoxels(images_rgb, images_thermal, poses)
 loss = Loss(rgb, temperature, rgb_gt, temperature_gt)
 thermoxels.optimize(loss)

Voxel optimization

Voxel optimization

poses = COLMAP(images_rgb)
# optimize color and temperature on voxels
thermoxels = Voxels(scene_size)
while i <= max_iteration:
 rgb, temperature = thermoxels(images_rgb, images_thermal, poses)
 loss = Loss(rgb, temperature, rgb_gt, temperature_gt)
 thermoxels.optimize(loss)

Volumetric Mesh Extraction

poses = COLMAP(images_rgb)
thermoxels = Voxels(scene_size)
while i <= max_iteration:
 rgb, temperature = thermoxels(images_rgb, images_thermal, poses)
 loss = Loss(rgb, temperature, rgb_gt, temperature_gt)
 thermoxels.optimize(loss)
# Reconstruct thermal volumetric mesh
vol_mesh, surface_temperature = filtering(voxels, threshold)

Finite-Element Analysis

poses = COLMAP(images_rgb)
thermoxels = Voxels(scene_size)
while i <= max_iteration:
 rgb, temperature = thermoxels(images_rgb, images_thermal, poses)
 loss = Loss(rgb, temperature, rgb_gt, temperature_gt)
 thermoxels.optimize(loss)
vol_mesh, surface_temperature = filtering(voxels, threshold)
# Run FEA simulation
simulation = FEA(vol_mesh, surface_temperature, env_conditions)

Results

Reconstruction Metrics

Method Heated Water Cup Heated Water Kettle Melting Ice Cup Building (spring) Building (winter) Double robot Exhibition Building Dorm 1 Dorm 2
ThermoNeRF
Plenoxels_t
Thermoxels
Figure 1: PSNR↑

Reconstruction Metrics

Method Heated Water Cup Heated Water Kettle Melting Ice Cup Building (spring) Building (winter) Double robot Exhibition Building Dorm 1 Dorm 2
ThermoNeRF 32.05 34.04 32.24 26.63 28.75 30.75 33.79 34.10 29.94
Plenoxels_t 31.93 34.40 42.05 35.24 34.62 33.29 27.50 39.24 35.50
Thermoxels 29.78 30.37 19.99 19.84 22.56 24.16 18.78 10.79 13.85
Figure 2: PSNR↑
ThermoNeRF 0.92 0.94 0.98 0.92 0.88 0.95 0.97 0.96 0.95
Plenoxels_t 0.92 0.98 0.96 0.96 0.93 0.92 0.97 0.98 0.97
Thermoxels 0.83 0.92 0.88 0.90 0.82 0.75 0.86 0.60 0.72
Figure 3: SSIM↑
ThermoNeRF 2.10 2.76 1.57 1.88 0.66 0.91 0.31 0.38 0.75
Plenoxels_t 0.87 1.41 0.11 1.36 0.40 0.49 1.00 0.32 0.39
Thermoxels 0.99 1.15 1.74 5.29 1.27 0.74 1.29 8.87 4.21
Figure 4: MAE↓

Simulation-ready models

Thermoxels temperature as initial condition

Simulation

Simulation

Conclusion and Future Work

  • The first method able to build 3D models compatible with FEA using sparse RGB images.
  • Metrics do not capture mesh quality.
  • Future works:
    • Improve surface approximation.
    • Integrate material properties.

References

Chen, Q., Shu, S., & Bai, X. (2025). Thermal3D-GS: Physics-induced 3D gaussians for thermal infrared novel-view synthesis. In A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, & G. Varol (Eds.), Computer vision – ECCV 2024 (pp. 253–269). Springer Nature Switzerland.
Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., & Kanazawa, A. (2022). Plenoxels: Radiance fields without neural networks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5501–5510.
Hassan, M., Forest, F., Fink, O., & Mielle, M. (2025). ThermoNeRF: A multimodal neural radiance field for joint RGB-thermal novel view synthesis of building facades. Advanced Engineering Informatics, 65, 103345. https://doi.org/10.1016/j.aei.2025.103345