Paper Paper Code Code Dataset Dataset

alt text

Abstract

Thermography is especially valuable for the military and other users of surveillance cameras. Some recent methods based on Neural Radiance Fields (NeRF) are proposed to reconstruct the thermal scenes in 3D from a set of thermal and RGB images. However, unlike NeRF, 3D Gaussian splatting (3DGS) prevails due to its rapid training and real-time rendering. In this work, we propose ThermalGaussian, the first thermal 3DGS approach capable of rendering high-quality images in RGB and thermal modalities. We first calibrate the RGB camera and the thermal camera to ensure that both modalities are accurately aligned. Subsequently, we use the registered images to learn the multimodal 3D Gaussians. To prevent the overfitting of any single modality, we introduce several multimodal regularization constraints. We also develop smoothing constraints tailored to the physical characteristics of the thermal modality. Besides, we contribute a real-world dataset named RGBT-Scenes, captured by a hand-hold thermal-infrared camera, facilitating future research on thermal scene reconstruction. We conduct comprehensive experiments to show that ThermalGaussian achieves photorealistic rendering of thermal images and improves the rendering quality of RGB images. With the proposed multimodal regularization constraints, we also reduced the model’s storage cost by 90%. Our project page is at https://thermalgaussian.github.io/.

RGBT-Scenes Dataset

The following 10 scenes were selected for both qualitative and quantitative analysis in our paper.

Each scene in the RGBT-Scenes dataset is displayed
Scene RGB Thermal MSX Views Temp. Range
Dimsum Dimsum RGB Dimsum Thermal Dimsum MSX 134(train) 20(test) 23.1°C - 60.0°C
Daily Stuff Daily Stuff RGB Daily Stuff Thermal Daily Stuff MSX 68(train) 10(test) 17.5°C - 56.3°C
Electric Bicycle Electric Bicycle RGB Electric Bicycle Thermal Electric Bicycle MSX 42(train) 6(test) 14.5°C - 18.5°C
Roadblock Roadblock RGB Roadblock Thermal Roadblock MSX 62(train) 9(test) 22.0°C - 27.0°C
Truck Truck RGB Truck Thermal Truck MSX 64(train) 9(test) 30.6°C - 249.0°C
Rotary Kiln Rotary Kiln RGB Rotary Kiln Thermal Rotary Kiln MSX 92(train) 14(test) 5.0°C - 60.4°C
Building Building RGB Building Thermal Building MSX 238(train) 35(test) 15.0°C - 24.0°C
Iron ingot Iron ingot RGB Iron ingot Thermal Iron ingot MSX 53(train) 8(test) 38.0°C - 350.0°C
Parterre Parterre RGB Parterre Thermal Parterre MSX 57(train) 9(test) 19.5°C - 27.5°C
Landscape Landscape RGB Landscape Thermal Landscape MSX 90(train) 13(test) 16.0°C - 23.0°C

RGBT-Scenes-extend-Dataset

The following are the four scenes extended by RGBT-Scenes.

Each scene in the expanded RGBT-Scenes dataset is displayed
Scene RGB Thermal MSX Views Temp. Range
Glass Cup Dimsum RGB Dimsum Thermal Dimsum MSX 123(train) 18(test) 17.0°C- 36.6°C
Transmission Tower Daily Stuff RGB Daily Stuff Thermal Daily Stuff MSX 154(train) 23(test) -26.4°C- 23.7°C
Dark Scene Electric Bicycle RGB Electric Bicycle Thermal Electric Bicycle MSX 75(train) 11(test) 17.5°C- 21.6°C
Plant Equipment Roadblock RGB Roadblock Thermal Roadblock MSX 192(train) 28(test) 27.8°C- 54.9°C

Dynamic scene demo

We also provide four dynamic scenes demos for our upcoming work to progress.

RGB GIF Thermal GIF
RGB GIF Thermal GIF
RGB GIF Thermal GIF
RGB GIF Thermal GIF

Comparisons

We present qualitative thermal image comparisons between our method, previous approaches (3DGS and ThermoNerf) , and the corresponding ground truth images from test views. We also show the training results of the MSX images, which are easier to apply.

thermal comparison

We present qualitative RGB image comparisons between our method and 3DGS from test views.

rgb comparison

To verify the effectiveness of our multimodal regularization term, we compare our multimodal regularization γ with manual adjustment of the thermal constraint coefficient in a truck scene. The comparison shows that our multimodal regularization approach reduces the storage space of RGB and thermal modalities while maintaining high image quality. At the same time, we visually present the Gaussian distributions of the original 3DGS method and our method with multimodal regularization.

dynamic_loss
(a) MR (γ) vs. Fixed coefficient
point number
(b)Gaussian distributions. Left: 3DGS; Right: Ours(MSMG)+MR

Citation

@article{lu2024thermalgaussian,
  title     = {ThermalGaussian: Thermal 3D Gaussian Splatting},
  author    = {Lu, Rongfeng and Chen, Hangyu and Zhu, Zunjie and Qin, Yuhang and Lu, Ming   and Zhang, Le and Yan, Chenggang and Xue, Anke},
  journal   = {arXiv preprint arXiv:2409.07200},
  year      = {2024},
}