Christopher Morse


Pool2Ocean: Synthetic Data Generation for Underwater Object Detection Using CycleGAN

This research presents a generative approach for data augmentation to help mitigate the risks and costs associated with underwater image data collection. Our goal is to improve underwater robot-to-robot detections with a more robust set of training data. The proposed approach generates synthetic ocean images through the use of CycleGAN, a Generative Adversarial Network that learns to provide mappings between unpaired images in a source domain (i.e. a robot operating in a pool) and a target domain (i.e. a robot operating in an ocean). The real and "fake" ocean images are compiled into three datasets (real ocean images, fake ocean images, and both combined) to train three object detection models. Each trained model is evaluated on a new set of real ocean data, where precision and recall scores, along with the average precision (AP) across various detection thresholds, are calculated. Our results show that the model trained on the combined set of real and fake ocean images significantly outperforms the other models, across each of these metrics. Through these results, this approach demonstrates a safe, fast, inexpensive, and non-invasive method for data augmentation.

Video file