A Hybrid Framework for Uncertainty-Aware Depth Prediction in the Underwater Environment

Abstract

Depth estimation is an important task that can be used in many applications such as scene reconstruction, virtual and augmented-reality, scene understanding and autonomous robots. Although several methods exist and perform well above water, for the underwater environment this task is undermined by challenging water properties. In this paper, we propose a framework for obtaining partial depth map supervision using a Structure From Motion (SfM) framework, and then train a deep neural network for real-time depth map estimation. Additionally, we estimate the model's uncertainty in the predictions, allowing us to filter out uncertain predictions and improve performance in all evaluated metrics. By estimating the model's uncertainty, the accuracy score A1 improved from 92.0% to 94.5%, and the Root Mean Squared Error (RMSE) reduced from 0.120 to 0.107.