Single-image spacecraft pose estimation

Visual navigation · 2026

Shiqi Ren and Pengcheng Wang

Two rows of spacecraft images with reference and estimated wireframes. Each row shows the three best-ranked estimates for its lighting condition. — The three lowest-error frames in each lighting condition for the baseline model. In the first Lightbox image, the estimated attitude is within 0.13° of the reference and the position is within 0.33 cm.

From eleven landmarks to a pose

A single camera sees a flat image of a three-dimensional target. A solar panel edge can be easy to locate under diffuse light and almost disappear in glare. An error of a few pixels at one landmark can rotate the estimated spacecraft even when the rest of its outline looks right.

Our 4.82-million-parameter network predicts the image locations of eleven known spacecraft landmarks and estimates their uncertainty. A geometric solver uses their known 3D coordinates to recover position and orientation. In a follow-up experiment, we rotate the same image four ways, run all four views through the same network and map the detections back into the original image. The pose that best fits the combined observations is selected.

What changes on held-out images?

We adapted the model using labeled laboratory images and evaluated it on 2,860 held-out SPEED+ frames. The table compares one pass with four-rotation refinement using the same weights. The pose score combines rotation and relative translation errors. Lower values indicate a closer estimate.

Mean pose score on held-out laboratory images
Lighting	Frames	One pass	Refined
Lightbox	2,022	0.0634	0.0415
Sunlamp	838	0.0741	0.0540

Six laboratory frames compare the reference pose in green, the single-pass estimate in dashed orange and the refined estimate in cyan. — Examples selected for improvement in both position and orientation. In Sunlamp frame (f), refinement reduces the position error from 15.56 to 1.83 cm and the orientation error from 2.87° to 1.23°.

SPEED+ imagery by T. H. Park, M. Märtens, G. Lecuyer, D. Izzo and S. D’Amico. Images were cropped and annotated with pose overlays. Dataset record · CC BY-NC-SA 4.0