Transfer Learning in Underwater Operations

This paper investigates a method for reducing the reality gap that occurs when applying simulated data in training for vision-based operations in a subsea environment. The distinction in knowledge in the simulated and real domains is denoted the reality gap. The objective of the presented work is to adapt and test a method for transferring knowledge obtained in a simulated environment into the real environment. The main method in focus is the machine learning framework CycleGAN, mapping desired features in order to recreate environments. The overall goal is to enable a framework trained in a simulated environment to recognize the desired features when applied in the real world. The performance of the learning transfer is measured by the ability to recreate the different environments from new test data. The obtained results demonstrates that the CycleGAN framework is able to map features characteristic for an underwater environment presented with the unlabeled datasets. Evaluation metrics, such as Average precision (AP) or FCN-score can be used to further evaluate the results. Moreover, this requires labeled data, which provides additional development of the current datasets.