BONK-pose: 6D Pose Estimation for Nordelbe Boats

• BONK-pose is a 6D pose dataset that combines monocular RGB images with AIS data to produce full 3D bounding box annotations for marine vessels.
• The dataset comprises 3,753 images and a 2D detection set, using YOLOX-X with performance metrics like [email protected] = 0.805 to benchmark detection accuracy.
• An automatic fusion pipeline leveraging PnP-based calibration significantly reduces projection error, offering actionable insights for maritime surveillance and research.

Searching arXiv for the cited BONK-pose paper and closely related Nordelbe/marine-pose papers. Boats on Nordelbe Kehrwieder (BONK-pose) is a publicly available 6D pose estimation dataset for marine vessels created by fusing monocular RGB imagery with Automatic Identification System (AIS) data from a static shore-based scene on the Nordelbe (Kehrwieder) stretch of the River Elbe in Hamburg. The dataset was introduced together with an automatic annotation pipeline that combines vessel detections from monocular RGB frames with synchronized Class A AIS broadcasts to generate 3D bounding boxes representing vessel pose, with the stated goal of avoiding manual annotation of vessel corners or heights. BONK-pose comprises 3,753 images with 3D bounding box annotations for 3,829 vessel instances, and is accompanied by a 2D ship-detection set of 1,000 images from the same scene (Holst et al., 20 Aug 2025).

1. Dataset definition and contents

BONK-pose is defined as a 6D-pose dataset in which each vessel is annotated with a full 3D bounding box specified by its 8 corners. In the dataset description, 6D pose is given as the $(x,y,z)$ position of the vessel centroid together with $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$, where yaw is derived from AIS heading and roll/pitch are approximately zero. Camera intrinsics $K$ and extrinsics $[R|t]$ are provided for the static view (Holst et al., 20 Aug 2025).

The paired 2D detection set contains 1,000 images annotated with rectangular 2D boxes in five sub-classes: ship, ship_leaving_frame, ship_moored, ship_partial, and subvessel. Standard metrics named for the resource are Intersection-over-Union (IoU), mean Average Precision (mAP)@IoU$=0.5$ for 2D detection, and reprojection error for pose (Holst et al., 20 Aug 2025).

A concise summary of the two components is given below.

Component	Size	Annotation content
BONK-pose 6D set	3,753 images; 3,829 vessel instances	3D bounding boxes, centroid, dimensions, rotation, camera parameters
2D detection set	1,000 images	Rectangular 2D boxes in five vessel sub-classes

This composition places BONK-pose at the intersection of object detection and pose estimation. A plausible implication is that it supports both direct 6D estimation and hybrid pipelines in which 2D detection acts as a proposal stage for geometric pose recovery.

2. Scene configuration, sensing, and calibration

The acquisition setup uses a static shore-based camera overlooking the Nordelbe (Kehrwieder) stretch of the River Elbe. The image stream consists of synchronized monocular RGB frames, and the auxiliary sensor stream consists of Class A AIS broadcasts. Images are stored as JPG at $1920\times1080$ from one static camera viewpoint (Holst et al., 20 Aug 2025).

Calibration is split into intrinsic and extrinsic stages. The intrinsic matrix $K$ is obtained via checkerboard calibration. A small set of manually picked world-image keypoints then fixes the extrinsic parameters $[R|t]$ by Perspective-n-Point (PnP). In the PnP formulation, the full pinhole model is used:

$\bar x \;=\;K\;\Pi\;\big[R\mid t\big]\;X,\qquad \Pi=\begin{bmatrix}1&0&0&0\0&1&0&0\0&0&1&0\end{bmatrix}.$

Given 3D keypoints $X_k$ in world coordinates (EPSG:4978) and their pixel positions $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$0, extrinsics are estimated via EPnP and refined by nonlinear reprojection error (Holst et al., 20 Aug 2025).

The scene itself is technically significant because it is a real harbor-river environment rather than a controlled laboratory setup. Related maritime RGB pose-estimation work in the same Nordelbe Kehrwieder harbor reports mixed diffuse daylight, strong building-façade reflections, rippling Fresnel highlights, moving vessels, overhead cranes, and periodic shadows from the quay (Mateos, 2020). This suggests that BONK-pose inherits the photometric and geometric complexity typical of operational shore-based maritime vision.

3. Automatic AIS–RGB fusion pipeline

The annotation pipeline begins with vessel detection in image space. The detector used as backbone is YOLOX-X, described as a single-stage model with a Darknet-style neck and head, pretrained on the MS COCO “boat” class and requiring no maritime-specific fine-tuning. On the BONK detection split defined as ship $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$1 ship_leaving_frame $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$2 subvessel, YOLOX-X achieves $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$3, $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$4, and $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$5 (Holst et al., 20 Aug 2025).

The pipeline then associates detections with AIS messages. Each AIS message supplies a position $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$6 in EPSG:4978 and a heading used to form vessel coordinate axes $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$7, with $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$8 forward, $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$9 port, and $K$0 up. The AIS location and heading are projected into the image either by a homography $K$1 or by the calibrated $K$2 transformation (Holst et al., 20 Aug 2025).

A “water-level plane segment” is defined around the AIS position by four world-space corners $K$3:

$K$4

where $K$5, $K$6, $K$7, and $K$8 are the bow, stern, port, and starboard distances from the AIS antenna. AIS age is compensated by dead reckoning,

$K$9

after which $[R|t]$0 (Holst et al., 20 Aug 2025).

The projected plane segment is enclosed in a minimal rectangle $[R|t]$1. Matching between detector outputs $[R|t]$2 and AIS-based windows is then performed by bipartite minimum-cost graph matching. The edge cost $[R|t]$3 uses normalized center displacement $[R|t]$4, bottom-edge difference $[R|t]$5, width difference $[R|t]$6, and the detector score:

• $[R|t]$7,
• $[R|t]$8,
• $[R|t]$9,
• $=0.5$0, with $=0.5$1 ignored if $=0.5$2 touches the frame border (Holst et al., 20 Aug 2025).

After association, the method corrects the bottom plane segment by visual constraints. The 2D box edges are reverse-projected into three planes $=0.5$3 through the camera center $=0.5$4; signed distances of the world corners to these planes are used to derive a correction vector $=0.5$5 so that the projected bottom face lies exactly against the left, right, and bottom box edges. Vessel height is then estimated from the top-edge box constraint by projecting the top edge into a world plane $=0.5$6, computing $=0.5$7, and setting $=0.5$8. The final 3D box corners are

$=0.5$9

Throughout this pipeline, no human labels of vessel corners or heights are needed (Holst et al., 20 Aug 2025).

4. Coordinate transforms and projection accuracy

A central methodological comparison in BONK-pose concerns the transformation from world coordinates to image coordinates. The homography-based approach assumes the water surface is approximately a plane $1920\times1080$0, selects $1920\times1080$1 3D keypoints $1920\times1080$2 on or projected onto $1920\times1080$3 with image correspondences $1920\times1080$4, and solves for a $1920\times1080$5 homography $1920\times1080$6 by minimizing

$1920\times1080$7

The image mapping is then

$1920\times1080$8

By contrast, the recommended PnP-based method uses the full pinhole model and estimated camera extrinsics rather than a planar approximation (Holst et al., 20 Aug 2025).

The mean projection-error comparison over all views is reported as follows.

Transformation	MAE (px)	MAE / width [%]
PnP	12.22	0.48%
Homography (water)	61.88	2.42%
Homography (PCA plane)	59.96	2.34%
Homography [Gülsoylu 2024]	22.65	1.67%

These figures support the paper’s conclusion that the Perspective-n-Point method achieves a significantly lower projection error than the homography-based approaches used before (Holst et al., 20 Aug 2025). In operational terms, the reduction from 61.88 px or 59.96 px to 12.22 px materially improves alignment between AIS-derived world positions and image evidence, which is essential because the final 3D box geometry is constrained by 2D detections.

The transform comparison also clarifies a common misconception: AIS alone does not provide a ready-made visual annotation. The paper explicitly motivates fusion by citing equipment reliability, data manipulation, and transmission delays as limitations of relying purely on AIS for location information (Holst et al., 20 Aug 2025). BONK-pose addresses this by using AIS as a world-space prior and monocular RGB detections as an image-space constraint.

5. Evaluation, annotation quality, and failure modes

For 2D detection, YOLOX-X was selected for speed and accuracy, with $1920\times1080$9 on moving-vessel classes. Other detectors listed in the evaluation—YOLOv3, DETR, Deformable DETR, and Cascade R-CNN r50—range down to approximately $K$0–$K$1 (Holst et al., 20 Aug 2025).

For AIS-to-detection association, the reported sample comprises 500 images and 581 visible vessels, from which 410 matches are obtained, corresponding to 70.6% correct associations. Among these 410 matched vessels, 375 are labeled “Good,” 33 “OK,” and 2 “Bad”; the same counts correspond to 64.5% of the total and 86.4% of the matched vessels for the “Good” category. Primary failure modes are missing AIS data, wrong matching by the cost function or two-vessel units, and detector false positives (Holst et al., 20 Aug 2025).

The paper further reports a correlation between per-instance 2D IoU of the projected automatic 3D box against manual 2D ground truth and manual quality labels: “Good” boxes are mostly at IoU $K$2, whereas “OK/Bad” boxes are mostly below this level (Holst et al., 20 Aug 2025). This provides a practical quality proxy when dense 3D ground truth is unavailable.

Recommended evaluation protocols are also stated. Standard train/validation/test splits of the 3,753 6D images are advised. The suggested metrics are $K$3 for detection, reprojection error in pixels and ADD-S for 3D pose, and 2D IoU of projected 3D boxes versus detection boxes as a proxy when CAD models are not available (Holst et al., 20 Aug 2025).

These evaluation results delimit the current reliability of the automatic labels. A plausible implication is that BONK-pose is most directly suitable for training or benchmarking models that can tolerate some annotation noise, especially where projected-box consistency and reprojection error can be used to filter instances.

6. Formats, access, and research uses

The dataset is distributed through the project page at https://fabianholst.github.io/BONK-pose/. The 2D detection set uses COCO-style JSON. The 6D set uses a separate JSON format with entries of the form $K$6 together with camera intrinsics $K$4 and extrinsics $K$5 per image. Python loader scripts on GitHub parse the JSON into PyTorch or TensorFlow datasets (Holst et al., 20 Aug 2025).

The stated applications are training and evaluation of 6D-pose networks for maritime surveillance, collision-avoidance for USV or shore-radar settings, traffic analysis, port automation, benchmarking of object-to-AIS association methods, and dimension-mismatch detection for customs or safety compliance (Holst et al., 20 Aug 2025). Because the annotations are fusion-based rather than manually drawn, BONK-pose is also a concrete example of automated label generation in a domain where manual 3D annotation is especially costly.

The dataset has already been positioned as a target domain for related marine-orientation research. In work on stable yaw estimation from UAV and USV viewpoints, the BONK-pose setting is described as one in which boats appear at varying distances as small image regions in video streams, and the proposed HyperPosePDF-based temporal aggregation is explicitly distilled “with the goal of adapting it to BONK-pose” (Kiefer et al., 2023). This suggests that BONK-pose can function not only as a 6D-pose benchmark but also as a substrate for probabilistic orientation estimation under severe scale variation and temporal ambiguity.

Within the broader maritime-vision literature, BONK-pose occupies a distinct position. AprilTags3D demonstrates robust 6-DoF estimation in the Nordelbe Kehrwieder harbor by using dynamic two-plane fiducials and RGB-only PnP-based fusion under glare and reflection (Mateos, 2020). BONK-pose differs in that it addresses uninstrumented real vessels by fusing monocular RGB detections with AIS rather than placing fiducials on targets. The common reliance on calibrated pinhole geometry and harbor-deployed RGB sensing indicates a shared methodological core, while the annotation objective of BONK-pose makes it primarily a dataset-generation framework rather than a fiducial tracking system.