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

Dynamic scene reconstruction from thermal infrared imagery remains insufficiently studied due to several inherent challenges, including low texture, low contrast, and radiometric ambiguity. In this paper, we present Thermal4D, a novel framework for reconstructing high-fidelity dynamic 3D scenes using only thermal images, without requiring visible-light inputs or auxiliary sensors. Built upon the 3D Gaussian Splatting paradigm, the proposed method introduces two key components. First, a frequency-aware attention module, termed TherHiLo, is designed to disentangle structural features across different frequency bands. Second, a physics-inspired atmospheric transmission module (ATM) is developed to model radiometric distortions caused by thermal imaging conditions. Although the reconstruction pipeline takes 8-bit thermal video sequences as input, high-precision 14-bit thermal frames are further exploited in TherHiLo to enhance attention learning with richer radiometric information. In addition, feature-level supervision from pretrained DINOv2 models is incorporated to improve structural consistency. To facilitate systematic evaluation, we also construct MVTD, a new multi-view dynamic thermal dataset. Experimental results on the MVTD and TI-NSD benchmarks show that Thermal4D consistently outperforms existing methods in both dynamic and static scenes, providing an effective framework for physics-consistent dynamic thermal scene reconstruction.