Datasets:
CiGS: Conduction in Gaussian Splats
This dataset has the multi-view thermal and visible-light images, with camera poses, used in the paper CiGS: Conduction in Gaussian Splats for Interior Temperature Field from Thermal Images. CiGS reconstructs an object as Gaussian splats from visible-light images and reads its surface temperature from thermal images. It then solves steady-state heat conduction (the Laplace equation) for the interior temperature, using a boundary integral equation on points extracted from the splats.
The dataset covers every experiment in the paper: the 18 synthetic scenes of the main evaluation, the real capture of a heated plaster teapot, and the scenes of the three inverse problems (buried heating pipe, facade insulation, hidden crack). Besides the images and poses, it has the metadata that converts image values to temperatures. It does not include ground-truth meshes, temperature fields or evaluation points.
Contents
| folder | what | scenes | size |
|---|---|---|---|
synthetic/ |
Blender renders: thermal radiance (EXR) and visible light (PNG) from the same 340 cameras | 18 | 2.2 GB |
real/teapot/ |
FLIR thermal + visible capture of a teapot on a hot plate, COLMAP poses, object masks | 2 heating powers | 0.4 GB |
inverse/ |
renders for the inverse problems: buried-pipe, facade-insulation, hidden-crack |
3 problems | 0.14 GB |
manifest.json lists the file count and size of every scene.
Synthetic scenes
Each scene is a mesh whose surface temperature is set by a harmonic function (solution_func in metadata.json).
We use Blender to render it at 512 x 512 pixels, both as a thermal image and in visible light, from 300 training cameras on a sphere around the
object and from 10 validation and 30 test cameras above it. Folder names are the scene keys used in the code, and the suffix
names the harmonic function.
| paper name | group | folder |
|---|---|---|
| Armadillo | solids | armadillo-xz |
| Botijo | solids | bojito-x2_z2 |
| Bunny | solids | bunny-xz |
| Hand | solids | hand-xyz |
| Fertility | solids | mother-linear_z |
| Nefertiti | solids | nefertiti-linear_z |
| Spot | solids | spot-x2_z2 |
| Springer | solids | springer-linear_z |
| Genus-6 | thin structures | genus6-linear_z |
| Voronoi | thin structures | voronoi-linear_z |
| Rind | thin structures | rind-linear_z |
| V-bunny | thin structures | vbunny-linear_z |
| Kangaroo | thin structures | kangaroo-linear_z |
| Dino | thin structures | dino-linear_z |
| Bracket | industrial parts | hx11tex-yz |
| Clamp | industrial parts | hx25tex-x2_y2_z |
| Hinge | industrial parts | hx18tex-2z2_x2_y2 |
| Housing | industrial parts | hx12tex-x2_y2 |
<scene>/
transforms_{train,val,test}.json cameras (NeRF-synthetic convention, see below)
metadata.json temperature range and the harmonic function
{train,val,test}/radiance/NNNN.exr thermal image, float32 RGBA
{train,val,test}/rgb/NNNN.png visible-light image, 16-bit RGBA
Thermal images
The color channels hold the normalized emitted radiance r (the three agree up to render noise), and the alpha channel holds
the object's pixel coverage. The temperature in kelvin is:
T = (r * (Tmax^4 - Tmin^4) + Tmin^4) ** 0.25, Tmin, Tmax = metadata.json temperature.actual_min, actual_max (293 K, 500 K)
Cameras
transforms_*.json follow the NeRF-synthetic convention. transform_matrix is camera-to-world with OpenGL camera axes
(x right, y up, looking down -z), camera_angle_x is the horizontal field of view, and w and h are the image size.
Each frame points to its visible-light image in file_path (without the extension, which 3DGS-style readers add) and in
rgb_file_path, and to its thermal image in radiance_file_path.
Real capture: teapot
The object is a solid teapot cast in plaster and heated from below on a hot plate at two powers, 100 W (100W/) and
150 W (150W/). A FLIR thermal camera captured each heating power from 360 viewpoints over the upper hemisphere, and each
thermal image has a registered visible-light image. The images are undistorted, and all images of one capture share one
COLMAP reconstruction.
teapot/
decode_temperature.py thermal EXR -> kelvin (numpy + OpenCV)
{100W,150W}/
radiance/IR_NNNNN.exr thermal image, float32, 464 x 348, FLIR raw signal normalized over the session
rgb/IR_NNNNN.jpg visible-light image, registered to the thermal image
masks/IR_NNNNN.png teapot mask
sparse/0/ COLMAP model (PINHOLE cameras, one per image; images.bin, cameras.bin, points3D.bin/.ply)
metadata.json normalization range of the raw signal and FLIR object parameters per image
A FLIR camera's raw signal is linear in in-band radiance, so the T^4 decode of the synthetic scenes does not apply. These
EXRs decode with FLIR's band-Planck model instead, which takes each image's emissivity, distance, and reflected and
atmospheric temperatures (stored in metadata.json) and the camera's calibration constants. decode_temperature.py
implements it:
python decode_temperature.py 150W IR_13216
Inverse problems
Buried heating pipe: inverse/buried-pipe
The scene is a 2.88 x 5.52 m room (room.json) with underfloor heating. Its floor follows the construction of Shin et al.
[2015], a PE-X pipe in a cement-mortar screed. The room is rendered from 40 training and 4 test cameras in its corners.
{train,test}/rgb/ holds the visible-light images and {train,test}/thermal_gray/ the thermal images as 8-bit gray PNGs.
To read a temperature, decode the sRGB transfer curve to the linear value g; then T [deg C] = 20 + 11 g.
Facade insulation: inverse/facade-insulation
The scene is a brick wall with an internal timber-frame insulation layer and hidden defects, seen from the street by 100
training and 10 test cameras. The folders are laid out as in the pipe scene, but thermal_gray/ holds 16-bit gray PNGs.
Decode the sRGB transfer curve to the linear value g; then T [deg C] = T_HI g, with T_HI = 2.5 from facade.json,
which also gives the wall dimensions.
Hidden crack: inverse/hidden-crack
The part is a HexMe propeller with a crack under the surface of its hub. It has three folders:
rgb/: 300 + 10 + 30 visible-light views of the part with a natural surface texture, used for the reconstruction.thermal-cracked/: 30 thermal views of the cracked part.thermal-intact/: the same 30 views of the intact part, used as the reference.
The thermal views (transforms_thermal.json) reuse every tenth camera of rgb/transforms_train.json; source_indices lists
which ones. They decode like the synthetic scenes, with the 293 to 500 K range in metadata.json.
Sources and credits
The data are released under CC BY 4.0. The rendered objects keep the terms of their sources, so please credit those as well:
- Armadillo, Bunny, Hand, Nefertiti, Spot, Springer: Oded Stein's mesh repository (each asset's original author is listed there).
- Botijo, Fertility: the AIM@SHAPE Shape Repository.
- Genus-6, Rind, Voronoi, V-bunny, Kangaroo, Dino: the shapes of Mehta, Chandraker and Ramamoorthi, A Level Set Theory for Neural Implicit Evolution under Explicit Flows, ECCV 2022.
- Bracket, Clamp, Hinge, Housing and the propeller: Beaufort et al., Hex Me If You Can, Computer Graphics Forum 2022.
- The heating-pipe construction: Shin, Rhee, Ryu, Yeo and Kim, Design of radiant floor heating panel in view of floor surface temperatures, Building and Environment 92, 2015.
Citation
@article{liu2026cigs,
title = {{CiGS}: Conduction in Gaussian Splats for Interior Temperature Field from Thermal Images},
author = {Liu, Zhenyuan and Seshadri, Bharath and Miller, Bailey and Kopanas, George and Gkioulekas, Ioannis and Bickel, Bernd},
journal = {arXiv preprint arXiv:2611.20283},
year = {2026},
doi = {10.48550/arXiv.2611.20283}
}
- Downloads last month
- 152