PGU-Depth: an underwater self-supervised monocular depth estimation method based on physical guidance

Abstract Despite recent progress in underwater monocular depth estimation, the field still faces numerous complex challenges. The performance of existing methods remains constrained by factors such as light attenuation, scattering effects, and color distortion inherent in underwater environments. To address these issues, we propose PGU-Depth, a physically guided underwater monocular depth estimation method that leverages physically inverted transmission maps to enhance model performance. A dual-branch underwater physical information extraction module optimizes features derived from the underwater medium-transmission map. An underwater physical information aggregation module then fuses RGB feature maps with the physically inverted transmission features at the encoder stage. The experiment results validate the superiority of the proposed PGU-Depth model. Finally, an information fusion decoder module integrates the multiscale features extracted by the extraction module into the decoding process, mitigating information loss.