Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 30 of the 30 references it cites were found in a registry or an index; 17 of the 40 sources its claims cite were checked for support.

Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor

Open access
Ahmed, Jamal, Arshad, Awais, Bang, Hyochoong, Choi, Yoonhyuk · 2024 · International Journal of Aeronautical and Space Sciences

Abstract

The objective of the proposed work is to perform monocular vision-based relative 6-DOF pose estimation of the non-cooperative target spacecraft relative to the chaser satellite in rendezvous operations. In this work, the convolutional neural network (CNN) is replaced by the high-resolution transformer network to predict the feature points of the target satellite. The self-attention mechanism inside the transformer provides the advantage of overcoming the inadequacies of the translation equivariance, 2D neighborhood awareness, and long-range dependencies in CNN. First, the 3D model of the target satellite is reconstructed using the inverse direct linear transform (IDLT) method. Then, the pose estimation pipeline is developed with a learning-based image-processing subsystem and geometric optimization of the pose solver. The image-processing subsystem performs target localization using CNN-based architecture. Then, the key points detection network performs regression to predict 2D key points using the transformer-based network. Afterward, the predicted key points based on their confidence scores are projected onto the corresponding 3D points, and the pose value is computed using the efficient perspective-n-point method. The pose is refined using the non-linear iterative Gauss–Newton method. The proposed architecture is trained and tested on the spacecraft pose estimation dataset and it shows superior accuracy both in translation and rotation values. The architecture has shown robustness against the drastically changing clutter background and light conditions in the space images due to the self-attention mechanism. Moreover, this method consumes less computation resources by using fewer floating-point operations and trainable parameters with low input image resolution.

Overview

CoChat Summary

TL;DR

The paper proposes a monocular 6-DOF relative pose estimation pipeline for non-cooperative spacecraft that replaces CNN keypoint detection with a high-resolution transformer (HRFORMER), followed by EPnP and Gauss-Newton pose refinement. On SPEED/SPEED+ data, it reports mean translation error of 6.48 cm and mean rotation error of 1.52°, with lower FLOPs and parameters than a comparable CNN.

Key findings

  • The full pipeline achieved mean translation error \(E_t = 6.48\) cm, median \(E_t = 3.78\) cm, mean rotation error \(E_q = 1.52^\circ\), and mean pose error \(e_T = 0.0349\); EPnP alone gave \(E_t = 10\) cm, \(E_q = 2.22^\circ\), and \(e_T = 0.0525\).
  • Keypoint detection with HRFORMER reached AP50 = 98%, AP75 = 98%, AP50-95 = 94.2%, AR50-95 = 95.87%, and mean OKS = 98.5%; target localization reached AP50 = 98.8%, AP75 = 97.9%, AP50-95 = 93%, mean IOU = 95%, and median IOU = 96%.
  • The transformer keypoint network used 3.33G FLOPs and 7.73M parameters at 256 × 192 input, versus 7.68G FLOPs and 28.53M parameters for the CNN-based HRNET comparison.
  • The authors report better translation and rotation errors than [13,16,31], but higher pose error than Wang et al. [17] (0.0123) and Chen et al. [15] (0.0094), which used higher-resolution images and/or more computationally complex solvers.

Methods

  • Monocular vision pipeline: 3D wireframe reconstructed from 14 selected images with 13 manually annotated key features using inverse direct linear transform (IDLT); input images preprocessed/augmented from 1920 × 1200 to 256 × 192; target localization via Faster-RCNN with HRNET-w18 backbone; keypoint detection via HRFORMER transformer producing 13 heatmaps.
  • Pose estimation: predicted 2D keypoints with confidence > 70% and at least 6 keypoints are matched by index to 3D model points; EPnP gives an initial pose, then Gauss-Newton method refines it until pose error reaches \(10^{-10}\) or a set iteration limit.
  • Data and training: SPEED and SPEED+ datasets, TANGO spacecraft, grayscale images, target-camera range 3–45 m; test set made of 3000 random images selected from the SPEED training set; transformer trained with Adam, learning rate 0.000125, batch size 25 on an NVIDIA TITAN XP GPU, using MSE heatmap loss.

Limitations

  • Stated: the work uses a single-domain dataset, and real space environment images are unavailable for training deep learning models; a domain gap remains between synthetic/source and real/target images, and the transformer approach did not significantly change results for this issue.
  • Stated: keypoint detection is less accurate when the target is very near (4–5 m) due to minimal camera field of view, and in dark areas; some keypoints are invisible or undetected due to contrast, light conditions, or target orientation.
  • Evident from the text: the reported test set is drawn from the SPEED training set rather than the official test or real-image sets, so the evaluation may not fully reflect generalization to unseen or real space imagery.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s42405-023-00703-3

Cites

30 of 30 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Images from the SPEED dataset that shows a challenging space environment
Fig. 1 Images from the SPEED dataset that shows a challenging space environmentCC-BY · From the PDF
Fig. 2 Geometrical model of the PnP problem
Fig. 2 Geometrical model of the PnP problemCC-BY · From the PDF
Fig. 4 a 14 cameras view to the target body; b a 3D reconstructed model of the target satellite
Fig. 4 a 14 cameras view to the target body; b a 3D reconstructed model of the target satelliteCC-BY · From the PDF
Image 4
CC-BY · From the PDF
Fig. 5 Target localization architecture with a feature extractor, RPN, and object detector
Fig. 5 Target localization architecture with a feature extractor, RPN, and object detectorCC-BY · From the PDF
Fig. 6 HRFORMER architecture shows the four stages of the model. The transformer blocks are used in the areas highlighted with blue and the information exchange blocks are used at the end of each stage to share information with low-resolution layers. a Parallel self-attention windows; b FFN 3 × 3 depth-wise convolution block to share information with multiresolution layer
Fig. 6 HRFORMER architecture shows the four stages of the model. The transformer blocks are used in the areas highlighted with blue and the information exchange blocks are used at the end of each stage to share information with low-resolution layers. a Parallel self-attention windows; b FFN 3 × 3 depth-wise convolution block to share information with multiresolution layerCC-BY · From the PDF
Fig. 7 a Training loss of target localization network; b training accuracy of target localization network
Fig. 7 a Training loss of target localization network; b training accuracy of target localization networkCC-BY · From the PDF
Fig. 8 Bounded box average precision vs epochs
Fig. 8 Bounded box average precision vs epochsCC-BY · From the PDF
Image 9
CC-BY · From the PDF
Fig. 9 a Training loss to detect key points; b keypoints training accuracy
Fig. 9 a Training loss to detect key points; b keypoints training accuracyCC-BY · From the PDF
Fig. 10 a Average precision vs average recall; b average precision-recall vs epochs
Fig. 10 a Average precision vs average recall; b average precision-recall vs epochsCC-BY · From the PDF
Image 12
CC-BY · From the PDF
Fig. 12 Error distribution graphs of pose errors using EPnP method
Fig. 12 Error distribution graphs of pose errors using EPnP methodCC-BY · From the PDF
Fig. 13 Error distribution graphs of pose errors using EPnP + GNM method
Fig. 13 Error distribution graphs of pose errors using EPnP + GNM methodCC-BY · From the PDF
Fig. 14 The sample images from the test dataset show the output key points and axis directions. These are images after the final pose estimation. The pose values are shown in Table 6
Fig. 14 The sample images from the test dataset show the output key points and axis directions. These are images after the final pose estimation. The pose values are shown in Table 6CC-BY · From the PDF
Image 16
CC-BY · From the PDF
Image 17
CC-BY · From the PDF
Fig. 15 3D-2D one-one linear mapping of key points a learning-based based keypoints network output b conventional method key points output
Fig. 15 3D-2D one-one linear mapping of key points a learning-based based keypoints network output b conventional method key points outputCC-BY · From the PDF

Verification

Claims the paper makes

15 Valid · 12 Not valid · 16 Not checked · 70 Own results and interpretations

44 more claims

Create a free account to see all 113 claims and to verify the 8 references still open.

See all claims — free account
  • Probably not a substantial claim. 32
  • It may state two separate claims as one. 39

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 28 references

Research works printed but not linked: 1

  1. Pauly L, Rharbaoui W, Shneider C, Rathinam A, Gaudillière V, Aouada D (2023) A survey on deep learning-based monocular spacecraft pose estimation: current state, limitations and prospects. Acta Astronaut 212:339–360. https://doi.org/10.1016/j.actaastro.2023.08.001 · DOI · Paper page

Cited by

See all 11

Metadata

Authors
Ahmed, J., Arshad, A., Bang, H., & Choi, Y.
Affiliations
Korea Advanced Institute of Science and Technology (KR)
Corresponding author
Bang, Hyochoong
Venue
International Journal of Aeronautical and Space Sciences
Published
2024-03-17
Volume
25
Issue
3
Pages
1146-1165
ISSN
  • 1229-9626 (ISSN-L)
  • 2093-274X
  • 2093-2480
OpenAlex
W4392890925
Cited by
11
References
28
Citations by year
2026: 6 · 2025: 4 · 2024: 1
FWCI
1.89
Type
journal article
Language
en
Open access
Licence
Locations
International Journal of Aeronautical and Space Sciences (publishedVersion, cc-by)
Funders
Korea Advanced Institute of Science and Technology
Topics
  • Physical Sciences › Engineering › Aerospace Engineering › Space Satellite Systems and Control
  • Physical Sciences › Engineering › Aerospace Engineering › Spacecraft Dynamics and Control
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
relative pose estimation, non-cooperative spacecraft, monocular vision, transformer models, self-attention, satellite rendezvous, keypoint detection, 3D reconstruction, perspective-n-point

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor. https://app.cochat.ai/papers/10.1007/s42405-023-00703-3
The paper
Ahmed, J., Arshad, A., Bang, H., & Choi, Y. (2024). Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor. International Journal of Aeronautical and Space Sciences. https://doi.org/10.1007/s42405-023-00703-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.