Project
State of the art deep learning models have shown great improvements in recent years but are still struggling with large domain shifts when deployed in new settings. This is particularly problematic in underwater environments as they display large variations based on geographic location, depth, weather and lighting conditions. At the same time, collecting data underwater comes with many practical limitations, and annotations are hard to acquire due to the often low footage quality and the need for expert input.
The goal of this PhD project is to study model generalization and domain adaptation approaches and improve them to fit the needs of the challenging underwater environments. Particular emphasis is given on low visibility settings with high turbidity where annotated data are especially scarce.
Funding
This project is funded by the Pioneer Center for Artificial Intelligence (P1)

Contact
PhD Fellow: Vasiliki Ismiroglou
Email: vasilikii@create.aau.dk
Supervisor: Thomas B. Moeslund
Email: tbm@create.aau.dk
Co-Supervisor: Malte Pedersen
Email: mape@create.aau.dk
Co-Supervisor: Stefan Hein Bengtson
Email: shbe@create.aau.dk