ThermalGate-GS: Frequency-Gated Graph Splatting for Thermal Novel View Synthesis

Area of Science:

• Computer Vision
• Thermal Imaging
• 3D Reconstruction

Background:

• Thermal infrared imaging is crucial for all-weather 3D perception.
• Heat conduction in thermal data acts as a low-pass filter, degrading high-frequency details and causing artifacts in 3D reconstruction.
• Existing methods face challenges in accurately modeling the diffusive nature of thermal fields.

Purpose of the Study:

• To develop a novel framework for high-fidelity 3D reconstruction from thermal infrared data.
• To address the limitations of standard 3D reconstruction pipelines in handling thermal information.
• To improve the accuracy and reduce artifacts in thermal 3D perception.

Main Methods:

• Proposed Frequency-Gated Graph Splatting (ThermalGate-GS) framework.
• Introduced a Frequency-Gated Anisotropic Diffusion mechanism to decouple thermal distributions and structural boundaries.
• Utilized high-frequency structural cues for adaptive gating weights to regulate thermal feature propagation.
• Regressed 3D Gaussian attributes using spectrally refined features.

Main Results:

• ThermalGate-GS significantly alleviates ambiguity in thermal 3D reconstruction.
• Achieved state-of-the-art performance on the ThermoScenes benchmark.
• Demonstrated a 7.94 dB PSNR improvement over prior physics-inspired baselines.
• Effectively suppressed cross-boundary bleeding while promoting surface smoothness.

Conclusions:

• ThermalGate-GS offers a robust solution for high-fidelity 3D reconstruction using thermal infrared imaging.
• The proposed frequency-gating and anisotropic diffusion approach effectively handles the challenges of thermal data.
• This method advances the capabilities of all-weather 3D perception systems.