BONK-pose

Boats On Norderelbe at Kehrwieder

A 3D bounding box Estimation Dataset for Marine Vessels

Overview

The Bonk-pose dataset is a novel, publicly available 3D bounding box estimation dataset for marine vessels, created by fusing monocular RGB images with Automatic Identification System (AIS) data. It addresses the gap in maritime datasets by providing fully annotated 3D bounding boxes for vessel pose estimation.

6D pose annotation showcase

Image 1
Image 2
Image 3
Image 4
Image 6
Image 7
Image 8
Image 9
Image 10
Image 11

Annotations

6D Pose Estimation

  • 3,753 images with 3D bounding box annotations including location, vessel dimensions and orientation.
  • 3,829 vessel annotations

Object Detection

  • 1,000 images with 2D bounding box annotations for vessel detection.
  • 1463 Vessels annotated in the foreground
  • 3957 Vessels and vessel like objects annotated

6D pose estimation

Created without human annotation cost

Annotations are created by an automated data fusion approach, making time-consuming human pose annotations unnecessary.
Annotations include vessel centroid location in the real world, vessel dimensions and the directions of the vessel coordinite system axes in the real world.
The intrinsic camera matrix is provided to project this data into the image.

Automatic Annotation Quality

86.4% of vessels labeled on par with human annotations.
94.5% of annotations are deemed of acceptable quality.

Object detection

Human annotations

Object detection bounding boxes were created by humans. Vessels are annotated in five classes. The classes allow differentiation between vessels in the foreground, moving vessels and vessels part of the background.

Object classes

  • Ship – Full vessels in the frame.
  • Ship Leaving Frame – Vessels partially out of the frame.
  • Ship Moored – Vessels anchored and stationary in the background.
  • Ship Partial – Partially visible or occluded vessels.
  • Subvessel – Independent units within a multi-vessel setup.
Performance of MSCOCO trained detectors on the object detection dataset.

Detector performance on vessels in the foreground

YOLOX-X

  • mAP@0.5:95: 0.626
  • mAP@0.5: 0.805
  • AR: 0.756

YOLOX-L

  • mAP@0.5:95: 0.603
  • mAP@0.5: 0.794
  • AR: 0.746

YOLOX-S

  • mAP@0.5:95: 0.542
  • mAP@0.5: 0.764
  • AR: 0.698

YOLOv3

  • mAP@0.5:95: 0.379
  • mAP@0.5: 0.659
  • AR: 0.549

DETR

  • mAP@0.5:95: 0.451
  • mAP@0.5: 0.696
  • AR: 0.645

Def. DETR

  • mAP@0.5:95: 0.510
  • mAP@0.5: 0.737
  • AR: 0.703

Def. DETR 2-Stage

  • mAP@0.5:95: 0.585
  • mAP@0.5: 0.775
  • AR: 0.761

Cascade R-CNN r50

  • mAP@0.5:95: 0.511
  • mAP@0.5: 0.725
  • AR: 0.680

Detector performance on All Vessel-like Objects

YOLOX-X

  • mAP@0.5:95: 0.398
  • mAP@0.5: 0.588
  • AR: 0.514

YOLOX-L

  • mAP@0.5:95: 0.363
  • mAP@0.5: 0.550
  • AR: 0.473

YOLOX-S

  • mAP@0.5:95: 0.277
  • mAP@0.5: 0.457
  • AR: 0.429

YOLOv3

  • mAP@0.5:95: 0.232
  • mAP@0.5: 0.476
  • AR: 0.332

DETR

  • mAP@0.5:95: 0.278
  • mAP@0.5: 0.513
  • AR: 0.398

Def. DETR

  • mAP@0.5:95: 0.337
  • mAP@0.5: 0.562
  • AR: 0.447

Def. DETR 2-Stage

  • mAP@0.5:95: 0.372
  • mAP@0.5: 0.571
  • AR: 0.531

Cascade R-CNN r50

  • mAP@0.5:95: 0.309
  • mAP@0.5: 0.527
  • AR: 0.427

Download and Links

Citation

@ARTICLE{holst2026fusing,
  author={Holst, Fabian and Gülsoylu, Emre and Frintrop, Simone},
  journal={IEEE Journal of Oceanic Engineering}, 
  title={Fusing Monocular RGB Images With AIS Data to Create a 3-D Bounding Box Estimation Data Set for Marine Vessels}, 
  year={2026},
  volume={51},
  number={3},
  pages={1676-1688},
  keywords={Marine vehicles;Estimation;Annotations;Modeling;Planing;Signal detection;Water;Object detection;Cameras;Distance measurement;3-D bounding box estimation;automatic identification system (AIS);data fusion;ship detection},
  doi={10.1109/JOE.2026.3695330}
}