UWDepth:Underwater image Depth estimation

In this paper, we propose a robust and efficient end-to-end model named UWDepth for fast monocular depth estimation that incorporates underwater domain knowledge into it as a supervised learning pipeline. We employ a new input space RMI for underwater domain-aware learning, including the red channel (R), the maximum of the green and blue channels (M) and the grayscale intensity (I). Since the attenuation constraint of the underwater light pillar (M) depends on the wavelength (M), the relative difference between the R and M channels embeds useful scene depth information in the I channel to preserve the structural content of the image. More importantly, visual results of experimental data show that its coarse pre-representation is quite accurate and can be used for 3D perception of low-cost underwater robots.