Abstract

Three-dimensional (3D) reconstruction of ships is crucial for maritime monitoring, enabling improved visualization, inspection, and decision-making in real-world maritime environments. However, most state-of-the-art methods rely on multi-view supervision, annotated 3D ground truth, or high computational cost, making them unsuitable for real-time applications at sea. In this work, we present an efficient pipeline for single-view 3D reconstruction of real ships using models trained entirely on synthetic data. At inference time, only a single real image is required. Our method leverages the Splatter Image network [1], which represents objects as sparse sets of 3D Gaussians for rapid reconstruction. We fine-tune this network on synthetic ShapeNet boats [2] and further refine it with a diverse set of custom-rendered ships to bridge the domain gap between synthetic and real data. The segmentation stage employs an enhanced ScatYOLOv8+CBAM model [3][4], with custom preprocessing to ensure alignment with the reconstruction pipeline. Final postprocessing steps include real-world metric scaling, orientation alignment, and placement on a geospatial map using AIS data and homography-based localization. The map visualization is powered by MapLibre GL JS [5]. Evaluation is performed on real images from the ShipSG dataset [6], captured in Bremerhaven, Germany. These examples, taken from the official validation split, confirm that the model—despite being trained solely on synthetic data—generalizes well to operational maritime conditions. The full system enables interactive 3D inspection of segmented ships in their real-world locations, without requiring any real-world 3D supervision. This pipeline provides a scalable, plug-and-play solution for future maritime situational awareness systems.

How to Use This Demo

• • Select a ship from the dropdown below. All examples use real images from the ShipSG dataset (Bremerhaven, Germany).
• • View the original input image, the result of segmentation using ScatYOLOv8+CBAM, and a video of the reconstruction process.
• • Explore the resulting 3D reconstruction as a geolocated point cloud overlaid on a real-world map.
• • On the map, zoom using the mouse scroll wheel. Rotate the 3D view by right-clicking and moving the mouse.

BibTeX

@inproceedings{carrillo2025ship3d,
  title={Synthetic-to-Real Domain Bridging for Single-View 3D Reconstruction of Ships for Maritime Monitoring},
  author={Carrillo-Perez, Borja and Sattler, Felix and Rodriguez, Angel Bueno and Stephan, Maurice and Barnes, Sarah},
  booktitle={SPIE Optics + Photonics},
  year={2025},
  organization={SPIE}
}