DSC++: Joint Optimization of Underwater 3-D Reconstruction and Sonar Poses via Differentiable Space Carving

Dense 3-D underwater reconstruction in turbid underwater environments is crucial for various offshore applications. Forward-looking imaging sonars are widely used in such conditions due to their robustness to visual degradation. While neural radiance field (NeRF)-based methods have shown promising results for underwater reconstruction, they typically assume known sonar poses—a strong assumption that rarely holds in real-world scenarios. Although jointly optimizing the poses together with the implicit model is widely used in vision-based methods, its application to sonar data is hindered by the lack of elevation information in sonar images, making direct pose estimation unreliable. To overcome this challenge, we introduce a frame pair filtering mechanism that selects images that facilitate joint optimization. Furthermore, we propose a point-cloud loss based on the sonar epipolar geometry to better constrain relative poses during the optimization process. Together, these components enable effective joint optimization of sonar poses and scene geometry. Additionally, we present an effective calibration method for the extrinsic parameters of the DVL-Sonar-IMU system, as a prerequisite for underwater reconstruction experiments using multimodal sensors. Extensive experiments conducted in both simulation and real-world environments have demonstrated the effectiveness and robustness of the proposed method.