The DOI matches its Crossref record; 65 of the 65 references it cites were found in a registry or an index; 13 of the 23 sources its claims cite were checked for support.
Domain Generalization for In-Orbit 6D Pose Estimation
Open accessAbstract
We address the problem of estimating the relative 6D pose, e.g., position and orientation, of a target spacecraft from a monocular image, a key capability for future autonomous rendezvous and proximity operations. Due to the difficulty of acquiring large sets of real images, spacecraft pose estimation networks are exclusively trained on synthetic ones. However, because those images do not capture the illumination conditions encountered in orbit, pose estimation networks face a domain gap problem, e.g., they do not generalize to real images. Our work introduces a method that bridges this domain gap. It relies on a novel, end-to-end, neural-based architecture as well as a novel learning strategy. This strategy improves the domain generalization abilities of the network through multitask learning and aggressive data augmentation policies, thereby enforcing the network to learn domain-invariant features. We demonstrate that our method effectively closes the domain gap, achieving state-of-the-art accuracy on the widespread SPEED+ dataset. Finally, ablation studies assess the impact of key components of our method on its generalization abilities.
Overview
CoChat Summary
TL;DR
The paper introduces a domain-generalization method for monocular 6D spacecraft pose estimation that trains only on synthetic images and combines a keypoint-to-pose network with histogram equalization, multi-task learning, and aggressive domain randomization. It reports state-of-the-art accuracy on the SPEED+ dataset, outperforming other domain-generalization methods except the more computationally intensive EagerNet.
Key findings
- On SPEED+ HIL test sets, the method achieves average errors of 9.3 cm and 4.3° on Lightbox and 14 cm and 6.9° on Sunlamp; median errors are 6 cm and 1.6° on Lightbox and 9 cm and 2.5° on Sunlamp, while synthetic validation medians are 3 cm and 1°.
- It outperforms all domain-generalization baselines except EagerNet, which is more computationally intensive, and is on par with most domain-adaptation approaches; the Lite-HRNet variant uses 1.2 GFlops versus 6.3 GFlops and achieves 0.14 m / 7.42° on Lightbox and 0.20 m / 15.12° on Sunlamp.
- Ablations show histogram equalization decreases errors by 25% on both test sets; adding both auxiliary segmentation tasks decreases errors by 40% on Lightbox and 45% on Sunlamp; RandAugment-style domain randomization reduces the pose score by 89% on Lightbox and 93% on Sunlamp, with texture and exposure augmentations contributing most.
- The continuous 6D rotation representation improves over quaternion: for the full PEM, S*P is 0.09 versus 0.15 on Lightbox and 0.14 versus 0.20 on Sunlamp.
Methods
- Design: a two-stage pipeline where a Keypoint Positioning Network predicts K = 11 keypoints from a 256×256 cropped image using an HRNet backbone, BiFPN, and DSNT, then a Pose Estimation Model uses a transformer encoder and MLP to regress normalized 3D translation and a continuous 6D rotation. Training uses histogram equalization, multi-task segmentation of the spacecraft and its six faces, and RandAugment-style domain randomization over six augmentation policies.
- Data: SPEED+ synthetic set of 59,960 grayscale 1920×1200 images plus 15,000 synthetic SPEED images, combined and split 80%–20% for training/validation; HIL test sets are Lightbox with 6,740 images and Sunlamp with 2,791 images. HIL images are used only for validation/test, and ROI cropping uses ground-truth keypoint locations.
- Analysis: evaluation uses the SPEED+ score S*P, average translation error ET, and average rotation error ER, with ablation studies on histogram equalization, multi-task learning, data augmentation policies, and PEM architecture.
Limitations
- Stated: HIL ground-truth positions and rotations are not perfectly accurate because of limited calibration accuracy; face segmentation masks are not available in SPEED+ and are approximated from keypoint coordinates, though the paper argues imperfect labels are acceptable for the auxiliary task.
- Stated: the method assumes a bounding box centered on the target is available, e.g., from an object detector or a previous pose estimate; in experiments, the ROI is cropped using ground-truth keypoint locations.
- Evident: evaluation uses HIL images rather than real on-orbit images, and the approach relies on a known CAD model and predefined keypoints for synthetic training.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 1 Not valid · 3 Not checked · 41 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“because those images do not capture the illumination conditions encountered in orbit”
- Also stated in: § 1 I ntroduction · p. 1 · hedged
“they are exclusively trained on synthetic images that hardly capture the illumination conditions encountered in Low-Earth Orbit and rely on a simplified spacecraft model.”
What the sources say
- Partly supportedcites [16] Bo Chen, Jiewei Cao, Alvaro Parra, and Tat-Jun Chin. Satellite pose estimation with deep landmark regression and nonlinear pose refinement. In Proceedings of the IEEE / CVF International Conference on Computer Vision Workshops , pages 0-0, 2019. doi: 10.1109 / ICCVW.2019. 00343. sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) reports a gap between synthetic and real spacecraft images and identifies challenging illumination as a problem. The Dataset and Metrics section (page unavailable) describes lightbox and sunlamp images designed to approximate Earth light and sunlight, but does not establish that synthetic images fail to capture illumination conditions encountered in orbit. (LLM, conf=0.82)
- Disputedcites [17] Sumant Sharma and Simone D'Amico. Pose estimation for non-cooperative rendezvous using neural networks. arXiv preprint arXiv:1906.09868 , 2019. doi: 10.48550 / arXiv. 1906.09868.Experimental — accuracy not yet measured
The SPACECRAFT POSE ESTIMATION DATASET (SPEED) section says the synthetic images’ solar illumination angles were specifically chosen to match those in actual Earth images. That contradicts the blanket claim that the synthetic images do not capture orbital illumination conditions; the paper does not claim they reproduce every such condition. (LLM, conf=0.84)
Checked by a reasoning model — may be wrong.
- Not checked — no full text of the sourcecites [18] Tae Ha Park, Sumant Sharma, and Simone D'Amico. Towards robust learning-based pose estimation of noncooperative spacecraft. arXiv preprint arXiv:1909.00392 , 2019. doi: 10.48550 / arXiv.1909.00392. source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1 I ntroduction · p. 1 · hedged
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“For example, in the SPEED + dataset [55], the Multi-Layer Insulation (MLI) represented in the CAD model is over-simplified compared to the actual one, as illustrated in Figure 4.”
What the sources say
- Partly supportedcites [55] Tae Ha Park, Marcus Märtens, Gurvan Lecuyer, Dario Izzo, and Simone D'Amico. Speed + : Next-generation dataset for spacecraft pose estimation across domain gap. In 2022 IEEE Aerospace Conference (AERO) , pages 1-15. IEEE, 2022. doi: 10.1109 / AERO53065.2022.9843439. sourceExperimental — accuracy not yet measured
The “On Realism of TRON Simulation” section says the spacecraft mockup was made from a CAD model, that some details were simplified or omitted, and that its body was covered in simulated MLI blankets. It does not say that the MLI representation in the CAD model was specifically over-simplified compared with the actual spacecraft. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“since Convolutional Neural Networks tend to overfit on the image texture [59]”
What the sources say
- Supportedcites [59] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 , 2018. doi: 10.48550 / arXiv. 1811.12231. source
Section 3.1 (p. 5) reports that the ImageNet-trained CNNs tested relied predominantly on texture rather than shape when the cues conflicted. Section 3.2 (p. 5) also reports poor transfer of ImageNet-trained features to images whose textures were made uninformative. These findings support the citing sentence’s paraphrase of a tendency to over-rely on texture. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Combining these augmentations through a RandAugment [64] policy significantly reduces the pose score, i.e ., by 89% on Lightbox and 93% on Sunlamp .”
What the sources say
- Supportedcites [64] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE / CVF conference on computer vision and pattern recognition workshops , pages 702-703, 2020. doi: 10.1109 / CVPRW50498.2020.00359. source
The citation identifies the RandAugment policy, not the source of the Lightbox and Sunlamp results. The Methods section describes RandAugment as selecting and combining transformations in a data-augmentation policy. The 89% and 93% reductions are presented as results of the citing paper, not attributed to this source. (LLM, conf=0.92)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“our texture-based augmentation achieves a reduction of 75% of the average pose score on both test sets.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
13 more claims
Create a free account to see all 53 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 30
- Probably not a substantial claim. 9
- The quote only partly states this claim. 3
Citations
Cites
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
An Algorithm for Least-Squares Estimation of Nonlinear Parameters
Marquardt, D. W. (1963) · Journal of the Society for Industrial and Applied Mathematics · doi:10.1137/0111030 · Cited by 30,728 · cited 1×
“The pose was then computed by solving a non-linear least squares problem using the Levenberg–Marquardt algorithm [21].”
MethodMobileNetV2: 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
Fischler, M. A., & Bolles, R. C. (1981) · Communications of the ACM · doi:10.1145/358669.358692 · Cited by 25,893
YOLOv3: An Incremental Improvement
Redmon, J., & Farhadi, A. (2018) · arXiv · doi:10.48550/arxiv.1804.02767 · Cited by 17,141
Advances in Neural Information Processing Systems 19
(2007) · The MIT Press eBooks · doi:10.52202/085713 · Cited by 15,539
An algorithm for the machine calculation of complex Fourier series
Cooley, J. W., & Tukey, J. W. (1965) · Mathematics of Computation · doi:10.1090/s0025-5718-1965-0178586-1 · Cited by 13,154 · cited 1×
“This Fourier-based augmentation is inspired by previous arts [52, 53] and consists in computing the Fast Fourier Trans-form (FFT) [54] of the image and adding noise to its magnitude while preserving its phase.”
MethodEfficientDet: Scalable and Efficient Object Detection
Tan, M., Pang, R., & Le, Q. V. (2019) · arXiv · Cited by 9,841
NeRF: representing scenes as neural radiance fields for view synthesis
Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., & Ng, R. (2022) · Communications of the ACM · doi:10.1145/3503250 · Cited by 6,852
Deep High-Resolution Representation Learning for Human Pose Estimation
Sun, K., Xiao, B., Liu, D., & Wang, J. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00584 · Cited by 5,984
Research works printed but not linked: 2
- VenturaJ. “Autonomous Proximity Operations for Noncooperative Space Targets,” Ph.D. Thesis, Technische Universität München, München, Germany, 2016.
- CassinisL. P. “Monocular Vision-Based Pose Estimation of Uncooperative Spacecraft,” Ph.D. Thesis, TU Delft, Delft, The Netherlands, 2022.
Cited by
Deep Learning-Based Object Pose Estimation: A Comprehensive Survey
Liu, J., Sun, W., Yang, H., Zeng, Z., Liu, C., Zheng, J., Liu, X., Rahmani, H., Sebe, N., & Mian, A. S. (2026) · International Journal of Computer Vision · doi:10.1007/s11263-025-02646-6 · Cited by 100
Feng, S., Dong-Yan, J., Cui, W., Wang, T., & Ma, M. (2025) · IEEE Sensors Journal · doi:10.1109/jsen.2025.3615727 · Cited by 8
“Legrand et al. [9] achieved texture enhancement by adding noise to the amplitude component while preserving the phase component after fast Fourier transform (FFT) processing.”
Online Test-Time Adaptation for On-Orbit Noncooperative Spacecraft Pose Estimation
Feng, S., Wang, T., Guo, Y., & Qiu, J. (2025) · IEEE Transactions on Aerospace and Electronic Systems · doi:10.1109/taes.2025.3623610 · Cited by 13
“As the PnP solver exhibits sensitivity to outliers, Legrand et al. [22] developed a pose estimation model that directly regresses poses from predicted keypoints.”
TAP-Track: Generalizable Spacecraft Pose Tracking by Tracking Any Points
Liu, Y., Qi, Z., Chen, P., Fan, H., Zhu, Y., Yang, J., & Zhang, Y. (2025) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2025.3584925 · Cited by 3
“…efforts toward bridging the domain gap between synthetic and real data have been made, via aggressive data augmentation on limited real data [31], introducing a learnable PnP [17], imposing geometric constraints during unsupervised domain adaptation [11], and incorporating additional data…”
Wang, Z., Liu, Y., & Zhang, E. (2024) · Aerospace · doi:10.3390/aerospace11110948 · Cited by 5
Metadata
- Authors
- Legrand, A., Detry, R., & De Vleeschouwer, C.
- Affiliations
- UCLouvain (BE)
- KU Leuven (BE)
- UCLouvain
- Venue
- Journal of Aerospace Information Systems
- Published
- 2025-11
- Volume
- 22
- Issue
- 11
- Pages
- 938-947
- ISSN
- 2327-3097 (ISSN-L)
- 1940-3151
- OpenAlex
- W4414268469
- Cited by
- 10
- References
- 28
- Citations by year
- 2026: 4 · 2025: 5
- FWCI
- 2.99
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Journal of Aerospace Information Systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Lirias (KU Leuven) (submittedVersion, other-oa)
- Digital Access to Libraries (submittedVersion, other-oa)
- arXiv (Cornell University)
- Funders
- Gouvernement Wallon
- Fonds De La Recherche Scientifique - FNRS
- Topics
- Physical Sciences › Engineering › Aerospace Engineering › Space Satellite Systems and Control
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Keywords
- 6D pose estimation, spacecraft pose estimation, domain generalization, domain gap, monocular images, autonomous rendezvous, proximity operations, multi-task learning, data augmentation, domain-invariant features
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: Domain Generalization for In-Orbit 6D Pose Estimation. https://app.cochat.ai/papers/10.2514/1.i011498
- The paper
- Legrand, A., Detry, R., & De Vleeschouwer, C. (2025). Domain Generalization for In-Orbit 6D Pose Estimation. Journal of Aerospace Information Systems. https://doi.org/10.2514/1.i011498
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.