Estimating Position and Orientation of Ships in Real Time Using Constrained Hand-Annotated Data

We propose a simple and easy-to-use pose estimation pipeline capable of estimating the position and orientation of ships at sea with respect to a camera. We implement an end-to-end pipeline: creating the datasets, annotating them in a format compatible for estimating pose, and finding the orientation angle of the ship at each frame. While existing techniques use standard pose estimation datasets, we use real world and synthetic data that we curated in-house. In this work, we use a real U.S. Coast Guard ship and a simulated naval ship as our objects of interest for pose detection. Built with a single stage keypoint detector as the backbone model, our method has the advantage over existing pose estimation techniques of being able to work with specific objects not found in common datasets. Our pipeline works with image or video datasets containing simple (and potentially hand-labeled) ground truth annotations, specifically keypoints overlaid on various features and bounding boxes placed around the ships. Our method's fast runtime speed and low angle error enables real-time inference in various real-world applications.