Sonar-to-RGB Image Translation for Diver Monitoring in Poor Visibility Environments

This work evaluates a method for generating visuallike images from sonar images using Generative Adversarial Networks (GANs), for the purpose of monitoring technical divers working in low visibility environments. The general goal is to enhance the interpretability of sonar images in order to assist emergency operators that monitor the safety of the divers. To train the models, sonar and visual data were collected over the course of three trials from two different sites, an indoor pool and a lake. We evaluate and compare two different generative models namely, a modified version of pix2pix and vid2vid. Results show that it is possible to recover visual information from sonar data when the camera image is highly disturbed.