Point-Supervised Seagrass Segmentation for 3D Underwater Habitat Mapping

Underwater imagery is often used in marine monitoring, for example in seagrass habitat mapping applications. It facilitates the collection of vital observations such as coverage estimation, species detection and recognition across diverse spatial scales. Analysing these images often relies on fully supervised segmentation methods, necessitating labour-intensive pixel-wise mask annotations and resulting in significant expert labelling costs. While previous efforts have explored weak supervision with point annotation, their focus has predominantly been on coarse seagrass segmentation, often limited to segmenting patches. This emphasis has overlooked the aspects of fine-grained segmentation and multi-dimensional mapping essential for seagrass habitat characterisation. In this paper, we propose a label-efficient fine-grained seagrass segmentation approach based on weak supervision with sparse point annotations, coupled with photogrammetric technology for underwater large-scale 3D habitat mapping. Our point-based supervision method, utilises 2D images to guide the segmentation process. We generate 3D mesh models and orthoimages of the seagrass habitat from collected underwater data. Subsequently, we map the semantic information directly into 3D space using a 2D-3D rasterisation-based approach and through a patch-based orthomosaic segmentation technique. The proposed approach is evaluated on the benchmark camouflaged object detection dataset, showing promising results in point-supervised segmentation. Additionally, it provides qualitative results in 2D segmentation, 3D seagrass mapping, and orthomosaic generation. Our approach offers a straightforward method to characterise seagrass habitats at a large scale, contributing to the advancement of marine conservation efforts.