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

The DOI matches its Crossref record; 67 of the 74 references it cites were found in a registry or an index; 3 of the 15 sources its claims cite were checked for support.

Satellite Pose Estimation Challenge: Dataset, Competition Design, and Results

Open access
Kisantal, Mate, Sharma, Sumant, Park, Tae Ha, Izzo, Dario, Martens, Marcus, D'Amico, Simone · 2020 · IEEE Transactions on Aerospace and Electronic Systems

Abstract

This paper describes the Satellite Pose Estimation Dataset (SPEED), the first publicly available machine learning set of synthetic and real spacecraft imagery of the Tango spacecraft. It assesses 48 competitor solutions for monocular pose estimation.

Overview

CoChat Summary

TL;DR

The Satellite Pose Estimation Challenge (SPEC) benchmarked monocular satellite pose estimation using the synthetic-and-real SPEED dataset, with 48 teams submitting results. The analysis found that PnP-based keypoint methods and a separate localization step improved accuracy, while Earth backgrounds and large target distances were the hardest conditions.

Key findings

  • 48 teams participated; top teams improved by two orders of magnitude over the baseline. UniAdelaide won on the synthetic test set with \(E_{syn}=0.0094\), \(e_q=0.41\pm1.50^\circ\), and \(e_t=0.032\pm0.095\) m; EPFL_cvlab achieved the best real-image score (\(E_{real}=0.1140\)).
  • PnP-based solutions significantly outperformed direct pose estimation: average orientation errors were \(9.76^\circ\pm18.51^\circ\) for direct methods versus \(1.31^\circ\pm2.24^\circ\) for PnP methods, and relative position errors were \(0.0328\pm0.0430\) m versus \(0.0083\pm0.0269\) m.
  • A separate localization step improved orientation accuracy: mean orientation error was \(29.66^\circ\pm46.10^\circ\) with separate localization versus \(48.03^\circ\pm49.38^\circ\) for combined localization-and-pose architectures, while position error distributions were nearly identical.
  • The most difficult samples had Earth background and small apparent target size due to large inter-spacecraft distance; black-background images were easier. The “super pose estimator” across all submissions had average orientation error \(0.34^\circ\pm0.38^\circ\) and average position error \(0.09\pm0.09\) m.

Methods

  • SPEC was hosted on the ESA Kelvins platform and used the Spacecraft Pose Estimation Dataset (SPEED): synthetic images generated with an OpenGL-based optical stimulator and real images captured with the TRON facility using a Tango spacecraft mockup. The dataset contained almost 15,000 synthetic and 300 real images, partitioned into 12,000 synthetic + 5 real training images and 2,998 synthetic + 300 real test images; all images were grayscale at \(1920\times1200\) pixels.
  • The competition metric combined orientation angular error and normalized position error into a single pose error, averaged over the test set. Submissions were primarily ranked on the synthetic test set and also evaluated on real images; during the competition only 20% of the test set was used for leaderboard evaluation.
  • The paper analyzes submissions from 48 teams and a post-competition survey answered by 20 teams, including the top 13. It compares top submissions, top 4 approaches, PnP versus direct pose estimation, the effect of separate localization, and sample difficulty using the best prediction across submissions for each test image.

Limitations

  • The paper states that real images are scarce, with only five labeled real training images, and that the real-image pose distribution is constrained by physical limitations of the TRON facility. It also notes discrepancies between synthetic and real images in texture, illumination, and eclipse of spacecraft features, and that real-image performance was weaker for most teams due to the domain gap.
  • The paper states that the competition metric’s balance between position and orientation sensitivity depends on the test set’s distance distribution, and that only 20% of the test set was used during the competition to prevent overfitting.
  • Evident from the text: the survey covered only 20 of 48 teams, so the approach comparison may not represent all submissions; the real test set contains 300 images and only five labeled real training images, limiting conclusions about generalization to real spaceborne imagery.

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

Integrity signals

Registry

Crossref · 10.1109/taes.2020.2989063

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
From the PDF
Fig. 2 Cropped versions of (a) the flight imagery captured during the PRISMA mission [26], (b) synthetic imagery in Beierle and D'Amico [43], (c) SPEED synthetic imagery, and (d) histogram comparison of image pixel intensities of the three images. They are cropped from the downscaled 224 × 224 pixel images.
Fig. 2 Cropped versions of (a) the flight imagery captured during the PRISMA mission [26], (b) synthetic imagery in Beierle and D'Amico [43], (c) SPEED synthetic imagery, and (d) histogram comparison of image pixel intensities of the three images. They are cropped from the downscaled 224 × 224 pixel images.From the PDF
Fig. 3 Left : TRON facility at SLAB. Right : Two examples of real training images from SPEED.
Fig. 3 Left : TRON facility at SLAB. Right : Two examples of real training images from SPEED.From the PDF
Fig. 4 Left : SPEED synthetic imagery. Right : SPEED real imagery.
Fig. 4 Left : SPEED synthetic imagery. Right : SPEED real imagery.From the PDF
Image 5
From the PDF
Fig. 5 Definition of spacecraft body reference frame ( B ), camera reference frame ( C ), relative position ( t BC ), and relative orientation ( R BC ).
Fig. 5 Definition of spacecraft body reference frame ( B ), camera reference frame ( C ), relative position ( t BC ), and relative orientation ( R BC ).From the PDF
Fig. 6 Position distributions of the pose labels across the dataset in the camera frame ( C ), for synthetic ( left ) and real ( right ) samples.
Fig. 6 Position distributions of the pose labels across the dataset in the camera frame ( C ), for synthetic ( left ) and real ( right ) samples.From the PDF
Fig. 7 Camera poses for real images in the Tango's body frame ( B ) from two views. The simplified wireframe model of the satellite is plotted in green, camera poses are plotted in red and black for test and training samples, respectively.
Fig. 7 Camera poses for real images in the Tango's body frame ( B ) from two views. The simplified wireframe model of the satellite is plotted in green, camera poses are plotted in red and black for test and training samples, respectively.From the PDF
Fig. 8 Distribution of the camera's relative positions of the synthetic images in the Tango's body frame ( B ) from two views. The satellite is in the origin, training and test camera poses are plotted in red and black for test and training samples respectively.
Fig. 8 Distribution of the camera's relative positions of the synthetic images in the Tango's body frame ( B ) from two views. The satellite is in the origin, training and test camera poses are plotted in red and black for test and training samples respectively.From the PDF
Fig. 9 Final results on the synthetic and real test sets.
Fig. 9 Final results on the synthetic and real test sets.From the PDF
Fig. 10 Position ( left ) and orientation ( right ) error distributions for direct and PnP solver based methods.
Fig. 10 Position ( left ) and orientation ( right ) error distributions for direct and PnP solver based methods.From the PDF
Fig. 11 Position ( left ) and orientation ( right ) error distributions highlighting the effect of a localization step prior to orientation estimation.
Fig. 11 Position ( left ) and orientation ( right ) error distributions highlighting the effect of a localization step prior to orientation estimation.From the PDF
Fig. 12 Top : Test images ranked by difficulty, measured as minimum pose error across all submissions. Bottom : Nine example images from different parts of the distribution. Images are shown with scaled colors to maximize contrast.
Fig. 12 Top : Test images ranked by difficulty, measured as minimum pose error across all submissions. Bottom : Nine example images from different parts of the distribution. Images are shown with scaled colors to maximize contrast.From the PDF
Fig. 13 Distribution of the minimum pose error in logarithmic scale with respect to the inter-spacecraft distance. The minimum pose error is computed across all submissions. Mean and standard deviation are calculated over one meter wide distance bins.
Fig. 13 Distribution of the minimum pose error in logarithmic scale with respect to the inter-spacecraft distance. The minimum pose error is computed across all submissions. Mean and standard deviation are calculated over one meter wide distance bins.From the PDF
Image 15
From the PDF
Image 16
From the PDF
Image 17
From the PDF
Image 18
From the PDF
Image 19
From the PDF
Image 20
From the PDF

Verification

Claims the paper makes

2 Valid · 3 Not valid · 9 Not checked · 70 Own results and interpretations

32 more claims

Create a free account to see all 84 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 26
  • It may state two separate claims as one. 24
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

8 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

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.

2 of 2

Citations

Cites

  • Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 2×

    “Various architectures were used from well known pre-trained models, such as ResNets [46], Inception v3 [47], and YOLO [48], [49], to custom models trained from scratch.”
  • ImageNet classification with deep convolutional neural networks

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 3×

    “It was made possible by increasing computational resources represented by the graphical processing units (GPU) and the availability of large-scale datasets to train the DNN, such as ImageNet for classification [5], MS COCO for object detection [28], and LINEMOD for pose estimation [29] of ordinary household objects.”
  • ImageNet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 1×

    “Starting with the success of AlexNet [5] in the ImageNet Large Scale Visual Recognition Challenge [6] in 2012, deep learning models have been outperforming traditional approaches on a number of computer vision problems.”
  • Distinctive Image Features from Scale-Invariant Keypoints

    Lowe, D. G. (2004) · International Journal of Computer Vision · doi:10.1023/b:visi.0000029664.99615.94 · Cited by 55,589 · cited 2×

    “int/ as SIFT [21], SURF [22], and ORB features [23].”
  • Microsoft COCO: Common Objects in Context

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · cited 1×

    “It was made possible by increasing computational resources represented by the graphical processing units (GPU) and the availability of large-scale datasets to train the DNN, such as ImageNet for classification [5], MS COCO for object detection [28], and LINEMOD for pose estimation [29] of ordinary household objects.”
  • Rethinking the Inception Architecture for Computer Vision

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138 · cited 1×

    “Various architectures were used from well known pre-trained models, such as ResNets [46], Inception v3 [47], and YOLO [48], [49], to custom models trained from scratch.”
  • A Computational Approach to Edge Detection

    Canny, J. (1986) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.1986.4767851 · Cited by 30,004 · cited 1×

    “oach to monocular-based pose estimation of a target spacecraft [6]–[12] would first extract hand-crafted features of the target from a 2D image. These features include Harris corners [13], Canny edges [14], lines via Hough transform [15], or scale-invariant features such as SIFT [16], SURF [17], and ORB features [18].…”
  • MobileNetV2: Inverted Residuals and Linear Bottlenecks

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00474 · Cited by 28,755 · cited 1×

    “However, the architectures exploit depthwise separable convolution operations [55] to significantly reduce the number of parameters associated with each network.”
  • YOLO9000: Better, Faster, Stronger

    Redmon, J., & Farhadi, A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.690 · Cited by 19,892 · cited 3×

    “Various architectures were used from well-known pretrained models, such as ResNets [46], Inception v3 [56], and YOLO [53], [54], to custom models trained from scratch.”

Show all 74 references

Research works printed but not linked: 3

  1. Sharma, Reduced-dynamics pose estimation for non-cooperative spacecraft rendezvous using monocular vision
  2. 10.1109/iccvw.2017.254 · doi:10.1109/iccvw.2017.254 · Paper page
  3. 10.1109/aero.2018.8396425 · DOI · Paper page

Cited by

See all 304

Metadata

Authors
Kisantal, M., Sharma, S., Park, T. H., Izzo, D., Martens, M., & D'Amico, S.
Affiliations
  • Stanford University (US)
  • European Space Agency (FR)
Venue
IEEE Transactions on Aerospace and Electronic Systems
Published
2020-10
Volume
56
Issue
5
Pages
4083-4098
ISSN
  • 0018-9251 (ISSN-L)
  • 1557-9603
  • 2371-9877
OpenAlex
W2988274083
Cited by
304
References
74
Citations by year
2026: 52 · 2025: 73 · 2024: 60 · 2023: 58 · 2022: 28 · 2021: 30 · 2020: 3
FWCI
17.86
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Aerospace and Electronic Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
  • Stanford Digital Repository
Funders
  • Air Force Office of Scientific Research
  • Defense University Research Instrumentation Program (FA9550-18-1-0492)
Topics
  • Physical Sciences › Engineering › Aerospace Engineering › Space Satellite Systems and Control
  • Physical Sciences › Engineering › Aerospace Engineering › Spacecraft Dynamics and Control
  • Social Sciences › Political Science and International Relations › Space exploration and regulation
Keywords
satellite pose estimation, monocular vision, on-orbit servicing, space debris removal, synthetic images, machine learning competition, performance indicators, testbed

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.

A language model drew the claims from the paper on 2026-10-12, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-12 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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: Satellite Pose Estimation Challenge: Dataset, Competition Design, and Results. https://app.cochat.ai/papers/10.1109/taes.2020.2989063
The paper
Kisantal, M., Sharma, S., Park, T. H., Izzo, D., Martens, M., & D'Amico, S. (2020). Satellite Pose Estimation Challenge: Dataset, Competition Design, and Results. IEEE Transactions on Aerospace and Electronic Systems. https://doi.org/10.1109/taes.2020.2989063

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.