Dense 3D semantic mapping in indoor scenes has been studied for several years and achieved reliable results. By applying such techniques in underwater scenes, the changes in the seafloor over time can be easily perceived. However, creating dense semantic maps for underwater images remains challenging due to the low quality of images, the lack of labeled data, and the often large datasets. The thesis proposes a large-scale, self-supervised framework to construct 3D dense semantic maps of the seafloor in order to generate a semantic changes detection map between surveys. The labeled images are rendered from the 3D reconstruction and are further used to train the 2D segmentation model. The 2D semantic predictions from multiple viewpoints will then be probabilistically fused into a map, producing a 3D semantic map.
The approach is evaluated with the accuracy of the 3D semantic segmentation, the consistency of 3D semantic map fussing from 2D segmentation, and the reliability of the change detection map.