WaveGS: Physics-Inspired Wavelet Splatting for Thermal Novel View Synthesis

ICLR 2026 Conference Submission16899 Authors

Keywords: 3DGS, Thermal Novel View Synthesis

Abstract: 3D thermal infrared reconstruction aims to reconstruct a three-dimensional model with thermal distribution information from multi-view thermal images or video sequences. Recent studies have shown that incorporating thermodynamic knowledge into 3D representations can achieve superior novel view synthesis performance. However, in the absence of temporal evolution and temperature calibration data, the representation learning becomes ill-posed. To address this problem, our key insight is to leverage the low-pass characteristic of heat conduction to model scene representations in the frequency domain. To this end, we first represent the 3D thermal field using a continuous vector-matrix (VM) decomposition, and parameterize the resulting factors with a learnable wavelet basis. This allows us to explicitly disentangle the scene representation into low-frequency components that capture smooth thermal variations and high-frequency subbands that encode structural details. Next, we devise a high-frequency masking strategy to suppress infrared noise while preserving salient details. Concurrently, this mask guides a learnable geometric deformation field to optimize geometric details by directly adjusting the anchor positions, thereby eliminating the need for explicit material parameters. Finally, the modulated wavelet coefficients are dynamically reconstructed into a spatial-domain feature field via a differentiable inverse wavelet transform. Extensive experiments on four datasets demonstrate that WaveGS consistently outperforms existing methods across multiple metrics.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16899