Thermal4D: Physics-Driven Gaussian Splatting for Dynamic Thermal Scene Reconstruction.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Thermal Imaging

Background:

• Dynamic 3D scene reconstruction from thermal infrared imagery is challenging due to low texture, low contrast, and radiometric ambiguity.
• Existing methods often require visible-light inputs or auxiliary sensors, limiting their applicability.

Purpose of the Study:

• To present Thermal4D, a novel framework for high-fidelity dynamic 3D scene reconstruction using only thermal images.
• To address the inherent challenges of thermal imagery, such as low texture and radiometric ambiguity.

Main Methods:

• The framework builds upon 3D Gaussian Splatting, incorporating a frequency-aware attention module (TherHiLo) and a physics-inspired atmospheric transmission module (ATM).
• It utilizes high-precision 14-bit thermal frames for enhanced attention learning and incorporates feature-level supervision from DINOv2 models for improved structural consistency.
• A new multi-view dynamic thermal dataset (MVTD) was constructed for systematic evaluation.

Main Results:

• Thermal4D demonstrates superior performance compared to existing methods on both dynamic and static scenes in benchmarks like MVTD and TI-NSD.
• The proposed TherHiLo module effectively disentangles structural features across different frequency bands.
• The ATM module accurately models radiometric distortions inherent in thermal imaging conditions.

Conclusions:

• Thermal4D provides an effective framework for physics-consistent dynamic thermal scene reconstruction.
• The method successfully reconstructs high-fidelity dynamic 3D scenes using only thermal imagery, overcoming previous limitations.
• This work advances the field of thermal infrared-based 3D reconstruction.