Practicality of Translation-Assisted Monocular Depth Estimation for Rainy Weather

Most monocular depth estimation models face difficulties under rainy conditions, and need to be robust for advanced driver-assistance systems. To overcome such issues, proposed works include changes in model architecture, along with further training. In this paper, we demonstrate the independent handling of rainy conditions from depth prediction with the use of our new dataset: NSTRain. Along with the unaltered versions, three sets of images are made: rain-only, neural-style-transferonly, and both. Furthermore, multiple rainfall intensities are created for each scene in the applicable sets for further analysis. Each set is used to calibrate a general adversarial network (GAN) model, and image-to-image translation is used to make depth model inputs with ideal weather and environment. Compared to the images prior to translation, outputs from the various trained GAN models have shown average improvements of up to 3.52 % in mean absolute relative error, 13.75 % in mean squared relative error, and 6.14 % in root mean squared error.