Geometry Aware 3D Multiview Thermal Reconstruction with Emissive Residual Decomposition Gaussian Splatting

Long-wave infrared (LWIR) imaging offers reliable perception in low-visibility scenes but suffers from sparse texture, causing floaters and poor geometry in current 3D thermal reconstruction methods. We cast multi-view thermal scene modeling as a thermal-specific volume rendering task and introduce a Gaussian-based representation designed for thermal data. We decomposes a scene into dense Emissive Gaussians for surface radiance and lightweight Residual Gaussians for floating artifacts, and augments them with a view-dependent emissivity network. Our thermal-only method outperforms prior thermal reconstruction pipelines on real world thermal and thermal-rgb novel-view synthesis (NVS) benchmark. Beyond high-quality NVS, our approach also recovers accurate thermal geometry and emissivity attributes, paving the way for dependable thermal perception in challenging environments.