Satellite Pose Estimation with Deep Landmark Regression and Nonlinear Pose Refinement
Open accessAbstract
We propose an approach to estimate the 6DOF pose of a satellite, relative to a canonical pose, from a single image. Such a problem is crucial in many space proximity operations, such as docking, debris removal, and inter-spacecraft communications. Our approach combines machine learning and geometric optimisation, by predicting the coordinates of a set of landmarks in the input image, associating the landmarks to their corresponding 3D points on an a priori reconstructed 3D model, then solving for the object pose using non-linear optimisation. Our approach is not only novel for this specific pose estimation task, which helps to further open up a relatively new domain for machine learning and computer vision, but it also demonstrates superior accuracy and won the first place in the recent Kelvins Pose Estimation Challenge organised by the European Space Agency (ESA).
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
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?
Even for SPEED and SPEED+, the only way to use a hybrid approach is to recover the 3D locations of keypoints via 3D reconstruction methods.
“Even for the standard datasets such as SPEED [19] and SPEED+ [20], the only way to use a hybrid approach is to recover the 3D locations of interested keypoints is via the 3D reconstruction methods [66].”
From A survey on deep learning-based monocular spacecraft pose estimation: Current state, limitations and prospects (2023) · cites this paper as [66]
Section 1.1 identifies SPEED as the dataset used, and Sections 3 and 3.1 describe reconstructing 3D landmarks by multi-view triangulation before estimating pose from 2D–3D correspondences. The paper supports this as its approach for SPEED, but does not establish that reconstruction is the only way to use a hybrid approach or address SPEED+. The indexed passages provide no page numbers. (LLM, conf=0.88)
Experimental — accuracy not yet measuredChen et al. won the Kelvins pose estimation challenge in 2020.
“They have won Kelvins pose estimation challenge in 2020.”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as [15]
The Introduction reports that Chen et al.’s method ranked first in the Kelvins Pose Estimation Challenge, and the Conclusion lists their team first. Neither passage dates that win to 2020; the cited paper is identified as a 2019 publication, so the year in the claim is unsupported by our copy. (LLM, conf=0.90)
Experimental — accuracy not yet measuredChen et al.'s method is unsuitable for space-grade hardware because of high image resolution and use of more floating-point operations.
“This work seems to be good for simulation but high image resolution and the use of more floatingpoint operations make this method unsuitable for space-grade hardware.”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as [15]
Section 3.3 (page unavailable) says the method increases its input window and output heatmaps from 256 × 256 to 768 × 768, supporting the high-resolution part. The paper does not report a comparison of floating-point operations or assess whether the method is unsuitable for space-grade hardware. (LLM, conf=0.84)
Experimental — accuracy not yet measuredThe transformer architecture uses almost half Flops compared to the CNN-based HRNET architecture used by [15] and [16].
“The information exchange between layers is only possible through convolution exchange blocks that isolate the transformer from other layers and that is the reason it uses almost half Flops as compared to the CNN-based HRNET architecture used by [15] and [16].”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as [15]
Section 3.3 (Landmark regression) confirms that the cited paper used HRNet for landmark prediction. The paper does not report transformer FLOPs, a near-half FLOPs comparison, or the proposed explanation involving convolution exchange blocks. (LLM, conf=0.85)
Experimental — accuracy not yet measuredChen et al. [15] mean rotation and translation errors are less than our work because they were participants of the KELVIN POSE estimation challenge and their focus is to perform well in the challenge.
“Chen et al. [15] mean rotation and translation errors are less than our work because they were participants of the KELVIN POSE estimation challenge and their focus is to perform well in the challenge.”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as Chen et al. [15]
The Introduction reports Chen et al.’s mean rotation and translation errors on the KPEC training set and says their test-set score ranked first in the challenge. The source does not compare those errors with the citing authors’ work or establish that challenge participation was the reason for any difference. (LLM, conf=0.84)
Experimental — accuracy not yet measured
Chen et al.'s method presents very low pose error.
“The image resolution of 768 × 768 is adopted and presents a very low pose error.”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as [15]
The mean IOU and median IOU for the bounded box subsystem are 95 and 96% respectively, almost similar to Chen et al. [15] and Massimo et al. [16].
“The mean IOU and median IOU for our bounded box subsystem are 95 and 96% respectively which is almost similar to the work presented by Chen et al. [15] and Massimo et al. [16].”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as Chen et al. [15]
Chen et al. used a higher image resolution of 768 × 768 and a simulated annealing pose solver to reduce the pose error.
“So, they have used a higher image resolution of 768 × 768 and a simulated annealing pose solver to reduce the pose error.”
From Transformer Network-Aided Relative Pose Estimation for Non-cooperative Spacecraft Using Vision Sensor (2024) · cites this paper as Chen et al. [15]
Previous satellite pose estimation approaches represent a satellite as a set of 2D keypoints and then estimate the satellite pose using PnP algorithms.
“Meanwhile, previous satellite pose estimation approaches [12], [15] represent a satellite as a set of 2D keypoints, and then estimate the satellite pose using the perspective-n-point (PnP) algorithms [27].”
From Bridging the Domain Gap in Satellite Pose Estimation: A Self-Training Approach Based on Geometrical Constraints (2024) · cites this paper as [12]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 32 · Cited by 203
Research works printed but not linked: 1
- [30]30. S. Sharma and S. D'Amico. Pose estimation for non- cooperative rendezvous using neural networks. In AAS/AIAA Astrodynamics Specialist Conference, 2019. 1, 2, 4, 6
- Authors
- Chen, B., Cao, J., Parra, A., & Chin, T.-J.
- Affiliations
- The University of Adelaide (AU)
- Venue
- 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)
- Publisher
- IEEE
- Published
- 2019-10
- Pages
- 2816-2824
- ISSN
- 2473-9936 (ISSN-L)
- OpenAlex
- W2970919703
- Cited by
- 203
- References
- 32
- Citations by year
- 2026: 16 · 2025: 40 · 2024: 47 · 2023: 45 · 2022: 26 · 2021: 19 · 2020: 10
- FWCI
- 24.97
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE International Conference on Computer Vision Workshops (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Adelaide Research & Scholarship (AR&S) (University of Adelaide) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Engineering › Aerospace Engineering › Space Satellite Systems and Control
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Engineering › Aerospace Engineering › Spacecraft Dynamics and Control
- Keywords
- satellite pose estimation, single-image pose estimation, geometry optimization, 3D model reconstruction, spacecraft docking, space debris removal
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.
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
