Towards an Open-Source Benchmark for Underwater Object Detection and Pose Estimation

Unmanned underwater vehicles (UUVs) are steadily expanding their merit in underwater inspection, maintenance, and repair tasks. A key step towards increasing their autonomy is to determine the pose of objects they interact with. This paper presents the work toward a benchmark dataset containing annotated poses and segmentation masks of industrial subsea handles and shackles. The latter are populated on a poster surrounded by fiducial markers to alleviate the object detection and pose estimation task. Video footage of the submerged poster is captured in a harbor by a remotely operating vehicle at varying depths. Detection and pose estimation of the handles and shackles are estimated by a region-based object-tracking approach. The dataset along with the tools to create it is publicly available at: https://github.com/OsloMet- OceanLab/UB_POSE