Attention-based Self-Supervised Hierarchical Semantic Segmentation for Underwater Imagery
Kurran Singh, Nick Rypkema, John J. Leonard · 2023
This work presents a method for semantic segmentation of underwater visual data. By leveraging recent advances in the machine learning community, this work shows that it is possible to overcome the lack of large labeled datasets for underwater scenes by using a self-supervised learning approach that can be pretrained on a large terrestrial dataset before being fine-tuned on a smaller underwater dataset. Results are provided demonstrating the effectiveness of this approach on a publicly available underwater dataset, as well as on experimentally collected data. Furthermore, this approach is shown to have a notable side effect in that multiple levels of detail for objects can be segmented out within a single forward pass through the network.