Marker free 6D pose estimation for underwater vehicles with monocular RGB cameras

Abstract

Unmanned underwater vehicles (UUV) are important for industry, science and explora- tion. A key step in increasing their autonomy and lowering costs is performing 6D pose estimation on objects it interacts with. This report details pose estimation methods and the process of creating a dataset for training neural networks. Footage captured in the sea using a ROV is annotated with segmentation masks and poses. This is made public along with the tools developed to create it. The dataset appears to be unique in its kind, similar datasets were made in pools and also not published. Further, the dataset is used to train Mask R-CNN for 2D object detection.