LiDAR-Based Semantic Segmentation for Marine Surroundings: Optimization strategies for segmentation classification in a marine environment

Abstract

In this thesis, methods for optimize an existing Convolutional Neural Network model for semantic segmentation are proposed. This is done through examining the size of the network, loss functions, dataset and how it can be preprocessed in different ways. The investigation show that preprocessing the data do not improve the model and that cross entropy loss is the best loss function when the dataset is highly imbalanced. The results from this project together with suggestions for future work shows bright results for future implementation.