With the increased occurrence of pollution in water bodies, there is a higher demand by the European Union for member countries to monitor their marine environments [1]. One of the reasons is that water contamination can be estimated by the behavioral change in animals living in the water body [2]. This creates a need for having a computer vision system that could monitor animals underwater. This system would highly depend on the underwater visibility that is influenced by several factors.
This master thesis builds on previous work by Pedersen et al. [3] and experiments with their unique underwater Brackish dataset which is highly influenced by turbidity. The aim of this project is to explore the effects of turbidity on the underwater object detector.
Firstly, the project uses real-time object detector YOLOv3 and manages to improve baseline results from the original paper by 9.2\% mAP. The result is verified by a newly introduced manually annotated dataset called the Brackish X dataset which can be used for evaluating the generality abilities of a trained model.
Secondly, this project evaluates which turbidity features are best for estimating turbidity.
Lastly, the best model is evaluated on the original test set divided into 3 subsets based on their estimated turbidity. The result of this experiment set a new course for a future experiment in a controlled environment.