Our article “Toward Training-Free Underwater 3-D Object Detection From Sonar Point Clouds: A Comparison of Traditional and Deep Learning Approaches” has been accepted for publication in the IEEEJournal of Oceanic Engineering. The paper compares traditional model-based and deep learning approaches for training-free underwater 3-D object detection from multibeam sonar point clouds. While the deep learning approach is trained on our synthetic sonar data, the traditional method uses template matching based on geometric object models. By evaluating both approaches on real bathymetry surveys from the Digital Ocean Lab the Baltic Sea, the paper contributes toward addressing the data scarcity problem in underwater perception, where annotated real-world sonar data is difficult and expensive to obtain.
Paper: https://ieeexplore.ieee.org/document/11506064
Journal of Oceanic Engineering: ieeexplore.ieee.org/xpl/RecentIssue.jsp