Thermal3D-GS: Physics-Induced 3D Gaussians for Thermal Infrared Novel-View Synthesis With a Large-Scale Dataset.

Area of Science:

• Computer Vision
• Thermal Imaging
• 3D Reconstruction

Background:

• Thermal infrared imaging offers all-weather, penetrating capabilities but faces challenges in novel-view synthesis.
• Existing methods struggle with coarse details and artifacts due to atmospheric transmission and thermal conduction.
• These limitations restrict accurate reconstruction of thermal scenes.

Purpose of the Study:

• To develop a novel-view synthesis method specifically for thermal infrared images.
• To address physical factors like atmospheric transmission and thermal conduction.
• To enhance reconstruction accuracy and detail in thermal scenes.

Main Methods:

• Introduced Thermal3D-GS, a physics-induced 3D Gaussian splatting method.
• Modeled atmospheric transmission and thermal conduction using neural networks.
• Incorporated sparse feature priors to improve reconstruction from sparse infrared data.
• Created the Thermal Infrared Novel-view Synthesis Dataset (TI-NSD) for validation.

Main Results:

• Thermal3D-GS achieved a 3.19 dB improvement in Peak Signal-to-Noise Ratio (PSNR) over baseline methods.
• The method effectively reduced floating artifacts and improved edge feature clarity.
• Experimental results validated the method's effectiveness on diverse thermal scenes.

Conclusions:

• Thermal3D-GS represents a significant advancement in thermal infrared novel-view synthesis.
• The physics-informed approach and sparse feature priors enhance reconstruction quality.
• The publicly released dataset and code facilitate further research in this domain.