In safety-critical environments such as firefighting, search and rescue, and industrial inspection, the presence of dense smoke severely hampers visual perception and degrades the performance of vision-based systems. Traditional dehazing and reconstruction methods are limited by their reliance on data-driven priors or assumptions of static, low-density smoke. We present SmokeSeer, a method that performs joint 3D scene reconstruction and smoke removal from multi-view video data, using both RGB and thermal imagery. Our approach leverages the reduced scattering properties of smoke in thermal images to recover robust spatial information.

Built upon the 3D Gaussian Splatting (3DGS) framework, our method explicitly decomposes a scene into smoke and non-smoke components. We validate SmokeSeer on synthetic and real-world datasets and release the first multi-view RGB-thermal smoke dataset. Our method is the first to integrate RGB and thermal modalities within a unified 3D reconstruction for dynamic, dense smoke environments.