3D Reconstruction of Underwater Scenes using Nonlinear Domain Projection

3D visual perception and scene reconstruction capabilities are paramount for autonomous robot navigation and intelligent transportation systems. Enabling these features are particularly challenging for underwater imagery due to optical distortion artifacts caused by light attenuation and absorption. These artifacts severely affect the classical 3D reconstruction pipelines because they solely rely on 2D feature matching and epipolar geometric constraints. In this paper, we propose a novel 3D reconstruction pipeline for underwater imagery that reduces the dependency on the standard 2D feature matching and ensures a denser scene recovery. It uses pixel-wise depth information as a guide to enhance underwater images as well as to generate sparse to semi-dense point clouds at a low computational cost. Extensive experiments demonstrate the feasibility and effectiveness of our approach in terms of reconstruction density and accuracy. It offers over three times faster inference than existing algorithms and standard software packages for multi-view 3D reconstruction.