Synthetic Data Generation for Deep Learning of Underwater Disparity Estimation
Elizabeth A. Olson · Corina Barbalata · Junming Zhang · K. Skinner · M. Johnson-Roberson
TL;DR
Real underwater images draped over randomized rendered terrain yield stereo training pairs with known depth, improving underwater disparity estimation over classical and terrestrial-trained methods.
BRIEF
Estimating depth from underwater stereo cameras is hard, and collecting ground-truth depth maps beneath the surface is costly. This work builds training data synthetically: real underwater images are projected onto 3D landscapes of randomized heights in a rendering framework, producing stereo image pairs with matching depth maps. A neural network trained on these pairs learns to match underwater image features. Reconstruction accuracy improved over traditional feature matching and over state-of-the-art networks trained on synthetic terrestrial data.
Abstract
In this paper, we present a new methodology to generate synthetic data for training a deep neural network (DNN) to estimate depth maps directly from stereo images of underwater scenes. The proposed method projects real underwater images onto landscapes of randomized heights in a 3D rendering framework. This procedure provides a synthetic stereo image pair and the corresponding depth map of the scene, which are used to train a disparity estimation DNN. Through this process, we learn to match the underwater feature space using supervised learning without the need to capture extensive real underwater depth maps for ground truth. In our results, we demonstrate improved accuracy of reconstruction compared to traditional computer vision feature matching methods and state-of-the-art DNNs trained on synthetic terrestrial data.