Pool2Ocean: Synthetic Data Generation for Underwater Object Detection Using CycleGAN
Abstract:In recent years, Autonomous Underwater Vehicles (AUVs) have been in the forefront of oceanic research and exploration. Since some of these AUVs depend on visual input to move, interact with the environment, and communicate, it is important for them to make accurate generalizations about their visual observations. To better equip these AUVs to perform well, it is necessary for them to have an object detection architecture that is sufficiently trained to safely operate in the real world. Unfortunately, collecting enough underwater imagery to properly train the detection model is often time consuming, expensive, and hazardous for both the humans and robots involved. Inspired by this problem, this research seeks to explore a more accessible method of data augmentation, through the intersection of deep learning and synthetic data generation, to improve underwater robot-to-robot detection.