Pseudo Forward Depth Estimation for Imaging Sonar Using Diffusion Models

Underwater robots often rely on sonar to perceive their surroundings. However, the absence of pitch angle information during sonar imaging hinders accurate perception of 3D environments. To address this issue, we propose SFDDiff, a diffusion model-based sonar pseudo-forward depth estimation paradigm for 3D reconstruction. In SFDDiff, we redefine the recovery of forward depth from sonar images as a generative denoising process. By leveraging a denoising U-Net within the pre-trained diffusion model, this method enables forward depth estimation with high performance under information-limited conditions. We validate the proposed method on two simulation datasets. Experimental results indicate that SFDDiff is capable of generating accurate and detail-rich forward depth maps. The code and dataset of this work will be available at https://github.com/sam-zyzhang/SFDDiff.git.