A large-scale, physically-based synthetic dataset for satellite pose estimation

Csaba Goldschmidt; K\'aroly Vass; Szabolcs Velkei

REVIEW 4 major objections 5 minor 19 references

A large-scale, physically-based synthetic dataset for satellite pose estimation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that a 640,000-image physically-based synthetic Hubble dataset can substantially close the simulation-to-reality gap for satellite pose estimation.

desk verdict A big, well-documented synthetic satellite dataset whose central domain-gap claim is explicitly untested; worth refereeing as a resource, not as a proof of transfer. read the letter →

arxiv 2506.12782 v1 pith:IADE3TUC submitted 2025-06-15 cs.CV

Szabolcs Velkei , Csaba Goldschmidt , Károly Vass This is my paper · ORCID

classification cs.CV

keywords satelliteposeestimationsyntheticdatasetdomainrandomizationphysically-basedrenderingHubbleSpaceTelescope6-DoFHDRimageryspacecraftproximityoperations

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents DLVS3, a synthetic data generator for training and testing 6-DoF satellite pose estimation, and introduces DLVS3-HST-V1, its first large release: 640,000 physically-based HDR images of the Hubble Space Telescope at 1024×1024 resolution, with ground-truth pose, 37 keypoints, semantic masks, depth, and normals. It claims that this combination of scale and physical modelling, including dynamic Sun/Earth/Moon lighting, secondary Earthshine illumination, procedurally aged materials, and per-image randomization, is a significant step toward closing the simulation-to-reality domain gap for autonomous proximity operations. The paper is explicit that real-world transfer remains a working hypothesis to be validated in future studies, and it positions the dataset as a benchmark for deep-learning pose estimators. A sympathetic reader would care because accurate pose estimation is a prerequisite for rendezvous, docking, and on-orbit servicing, and labelled real-space imagery is scarce. The full planned release grows to one million images with the addition of 320,000 articulated and 40,000 scenario-based images.

What carries the argument

The load-bearing mechanism is the DLVS3 image-generation pipeline: domain-randomized physically-based rendering, with a MaterialX-based material library that randomizes reflectivity, roughness, anisotropy, and aging effects such as UV discoloration, tarnishing, and micrometeorite damage in every image. A 360-degree HDR environment map rendered in Unreal Engine serves as an emissive dome around the satellite, so Earthshine contributes real secondary illumination during Houdini/Karma ray tracing. SPICE supplies astronomically consistent celestial positions, and a post-processing stage applies sRGB conversion, panchromatic filtering, vignetting, blooming, lens flares, and camera noise to the EXR layers. This pipeline is what carries the argument that synthetic imagery can be simultaneously large-scale, physically plausible, and richly annotated.

What would settle it

Train a pose estimator on DLVS3-HST-V1 and evaluate it on the SPEED+ real-image benchmark or on real HST mission photos; if accuracy is no better than training on older synthetic sets, the claim that this dataset closes the domain gap is refuted. A second check is radiometric: compare DLVS3 renderings to real HST imagery under matched Sun/Earth geometry; large photometric mismatches would undercut the physical basis for the transfer claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a production-grade simulator can generate space imagery that is both large and physically varied enough to make deep-learning pose estimators transfer to real operations. DLVS3 combines Unreal Engine 5 real-time rendering and Houdini/Karma offline ray tracing, a custom MaterialX material library with procedural aging and randomized surface variation, SPICE-based orbital geometry, and a 360-degree 32-bit floating-point HDR environment map that turns Earthshine into an active secondary light source. The generator emits multi-layer OpenEXR files carrying color, camera-space normals, depth, and 19-part cryptomatte segmentation in a single container, plus 2D projections of 37 keypoints. The released HST set follows true orbital trajectories from a fixed epoch, with 100 perturbed chaser and target positions per five-minute interval over 640 orbits. The intended payoff is that models see a broader distribution of lighting, surface states, and backgrounds than earlier synthetic sets, which the paper compares against at 5,000 to 180,000 images with simpler illumination models.

Load-bearing premise

The load-bearing premise is that randomized synthetic imagery will improve pose estimation on real spacecraft imagery; the paper itself labels this a working hypothesis intended to be validated in future studies, so if that transfer does not materialize the dataset's central value is unproven.

Editorial extensions

If this is right

Training pose networks on the initial 640,000 HDR images exposes them to lighting, surface, and background variation that prior public datasets, typically 50,000 to 100,000 images, do not cover.
Because Earthshine is rendered as a live secondary source, the dataset includes partially shadowed and Earth-lit configurations that earlier simulators with few direct light sources omit.
Procedural material aging and randomized multilayer-insulation wrinkling let the dataset span 25 to 30 years of unknown surface degradation without committing to one assumed state.
The planned 320,000-image articulated release will exceed what rigid-body PnP solvers can exploit, pushing the field toward pose estimators that handle variable joint configurations.
Metadata follows the real HST orbit from a fixed epoch via SPICE, so the same images can be used for trajectory-based and temporal pose estimation, not just per-frame training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the working hypothesis holds, DLVS3-HST-V1 could become a standard pretraining corpus, so the decisive test is not visual realism but downstream accuracy on real imagery such as the SPEED+ benchmark.
The multi-layer EXR format with depth and normals suggests a testable extension the paper does not perform: evaluating whether normal or depth supervision from this set improves pose accuracy versus color-only training.
The randomization strength is currently fixed; a natural extension would sweep randomization magnitude against real-world validation accuracy to find the operating point that best trades diversity against photorealism.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing to know: DLVS3-HST-V1 is a real resource—640k HDR images with keypoints, depth, normals, 19-part masks, SPICE-based orbital metadata, procedural material aging, and Earthshine lighting. That combination is new at this scale, and the planned 320k articulated images address something most prior benchmarks ignore. The paper describes the pipeline in enough detail that the dataset could be a useful training and benchmarking corpus for satellite pose estimation.

Second thing: the paper’s main advertised payoff—'closing the domain gap'—is not tested anywhere. There are no pose-estimation experiments, no training runs, no SPEED+ or real-image evaluation. Section 4.5 says explicitly that the material-randomization strategy is 'a working hypothesis... intended to be validated in future studies.' The stress-test note is right: the central utility claim rests on an untested assumption. That makes the contribution conditional, not accepted.

What I liked. The annotation schema is thorough (37 keypoints, segmentation, depth, normals), the metadata design is sensible, and the inclusion of Earthshine as an HDR dome is a genuinely useful idea. The Table 2 comparison is a fair overview. I also credit the authors for flagging that not every image is meant to be photoreal and that users should help validate the randomization.

Soft spots, in proportion. The missing experiments are the big one. Without them, the paper is a dataset description, and it should say so rather than claim a 'significant step.' The 'orders of magnitude larger' phrase overstates things relative to SPADES (179k) and URSO (150k); 640k–1M is several times larger, not orders of magnitude. Section 4.7 says each 1000-image subset is the daytime part of one orbit and the next sequence starts at the following dawn, so eclipse lighting is systematically excluded. That is a real sampling bias if the goal is operational robustness. Earthshine is called 'physically-based,' but no quantitative check is reported; visual plausibility is not validation. Finally, full data release is via request/SFTP, not open download, which will slow independent verification. None of this is disqualifying for the dataset’s existence; it is disqualifying for the paper’s overreach.

The citation pattern looks standard—SPEED, SPEED+, SHIRT, SPADES, SEENIC, SPARK, URSO and the main simulators are all covered. There are no mathematical claims to check; the load-bearing claims are empirical, and they are not yet supported.

Who this is for: people building or evaluating satellite pose estimators who want a large, varied synthetic corpus and are willing to treat it as a candidate training set. Not for readers who need evidence that synthetic-to-real transfer actually works.

Recommendation: send to peer review. A serious referee should ask for at least one transfer experiment on an existing benchmark, or a softened title and abstract plus a concrete validation roadmap. The dataset itself deserves scrutiny; the current framing does not.

Referee Report

4 major / 5 minor

Summary. The paper presents DLVS3, a synthetic dataset generation pipeline for satellite pose estimation, and introduces DLVS3-HST-V1, a dataset of 640,000 physically-based HDR images of the Hubble Space Telescope with annotations for 6-DoF pose, 37 keypoints, semantic segmentation, depth, and normals. The pipeline combines real-time (Unreal Engine) and offline (Houdini Karma) rendering, MaterialX-based procedurally aged materials, SPICE-driven orbital ephemerides, secondary illumination from Earthshine, and OpenCV post-processing for sensor artifacts. The authors claim the dataset is orders of magnitude larger and more diverse than previous benchmarks and positions DLVS3 as a significant step toward closing the simulation-to-reality domain gap for autonomous spacecraft operations.

Significance. If the dataset's utility were demonstrated, DLVS3-HST-V1 would be a substantial resource for the satellite pose estimation community: it provides a large number of HDR images with dense annotations, explicit modeling of Earthshine and material aging, and a planned articulated extension with 320,000 additional images. The paper's strengths include the use of SPICE for astronomically consistent geometry, the MaterialX-based procedural material system, the detailed metadata schema (transform matrices, keypoints, celestial body positions), and the availability of a 10,000-image demo subset for community evaluation. The central claim, however, is that this synthetic data helps close the domain gap; that claim is not supported by any experiment in the manuscript, and the paper itself labels the core randomization strategy as a 'working hypothesis' in Section 4.5.

major comments (4)

[Sections 5 and 4.5] The paper reports no pose estimation experiments, no domain-transfer evaluation, and no comparison against established benchmarks such as SPEED+ or SPADES. The conclusion that DLVS3 'provides a robust foundation' and 'positions DLVS3 as a significant step toward closing the domain gap' is therefore unsupported by direct evidence. Section 4.5 explicitly states that the material-randomization approach is 'a working hypothesis, which is intended to be validated in future studies.' For a dataset paper that makes domain-gap closure its central value proposition, at least one baseline experiment (e.g., training a simple keypoint or pose regression network on DLVS3 and evaluating on a real or test-bench dataset) is needed to substantiate the claim.
[Section 2 and Table 2] The claim that DLVS3 generates datasets 'orders of magnitude larger' than previous efforts is not supported by the paper's own comparison. Table 2 lists the largest prior datasets as SPADES (179,400 images) and URSO (150,000 images); DLVS3's initial release has 640,000 images and a planned total of 1,000,000. This is a factor of roughly 4–6, not an order of magnitude. The wording should be corrected to 'substantially larger' or similar, and the comparison should account for the number of annotated channels rather than raw image count.
[Section 4.7] The metadata description states that each 1000-image subset is 'the daytime portion of one orbital period' and that sequences are initiated at the following dawn. Since HST orbits in low Earth orbit and periodically passes through Earth's eclipse, restricting rendering to daytime portions systematically removes eclipse illumination conditions that occur in operational imagery. This selection bias affects the pose/lighting distribution and should be justified quantitatively (e.g., what fraction of real HST servicing imagery is taken in eclipse?) or remedied by including such conditions.
[Sections 3.5 and 3.2] The paper repeatedly uses the label 'physically-based' for the rendering and specifically for Earthshine (Section 3.5) and material aging (Section 3.2). However, no quantitative validation is provided that the secondary-illumination HDR dome produces physically correct radiance, nor that the procedurally aged materials match measured reflectance properties of real spacecraft surfaces. The claim rests on visual plausibility. The authors should either provide a radiometric or comparative validation (e.g., against known BRDF data or real HST photographs) or temper the wording to 'physically motivated' until such validation exists.

minor comments (5)

[Section 3.6] There is a typo: 'Poission' should be 'Poisson'.
[Section 4.7] The sentence 'the target is he observed satellite' contains a typo ('he' should be 'the').
[Table 2] The table rows are not consistently aligned with the column headers (e.g., 'Celestial position' and 'Depth map' appear to be in the column for DLVS3, but the preceding row 'Keypoints' has checks for only some datasets). Please reformat the table to make clear which properties apply to which dataset.
[Section 4.7 metadata list] The metadata field 'SunCenter (Vector2 1)' appears to include an extraneous '1'; it should likely be 'SunCenter (Vector2)'.
[References] Reference [12] (SurRender) contains a typo ('F ormation') and reference [7] is missing page numbers or a DOI. Additionally, the dataset portal URL in reference [19] should be supplemented with a persistent identifier if one becomes available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dataset paper contains no fitted predicate or derivation that reduces to its inputs, and the domain-gap claim is explicitly a working hypothesis for future validation.

full rationale

DLVS3 is a dataset-generation paper. It fits no parameters, derives no predictions from fitted constants, and does not invoke any author-uniqueness theorem. The central assertion that DLVS3-HST-V1 helps close the simulation-to-reality domain gap is presented as an untested working hypothesis ("This is considered a working hypothesis, which is intended to be validated in future studies", Section 4.5), so it is not passed off as a derived result. The only self-reference, the dataset portal [19], is an access link rather than load-bearing evidence. Statements about physical accuracy of materials and Earthshine are unvalidated rendering choices, which are evidential weaknesses (correctness risk), not circular reasoning. The "daytime portion of one orbital period" metadata description (Section 4.7) raises a potential sampling-bias concern, but it is a data-construction choice, not a derivation that reduces to its own output. Under the standard that circularity requires a quoted reduction or a fitted parameter renamed as a prediction, no circular step is present.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on modeling choices rather than fitted constants. The listed free parameters are dataset configuration values chosen by the authors; the axioms are domain assumptions about rendering fidelity, material realism, and ground-truth correctness that are not independently validated.

free parameters (6)

Camera field of view = 90 degrees
Fixed to an ideal pinhole, distortion-free camera; chosen by authors without sensitivity analysis.
Chaser-target distance range = 15-30 meters
Set based on expected proximity operations; no analysis of distance distribution effects.
Keypoint count = 37
Chosen by authors to capture HST structure; no ablation study.
Segmentation mask count = 19
Defined by authors for part-level segmentation; no justification provided.
Perturbed positions per time interval = 100
Arbitrary sampling density; no convergence analysis.
Dataset scale = 640,000 initial, 1,000,000 planned images
Scale chosen by authors; no sample-size or power analysis.

assumptions (6)

domain assumption HST 3D model is cm-level accurate
Used throughout generation; no source or accuracy certificate is given (Section 4 and Table 1).
domain assumption Rendering engines produce physically accurate images
Unreal Engine and Houdini are assumed to yield physically plausible reflectance and lighting (Section 3).
domain assumption Material randomization stays within physically plausible bounds
Explicitly stated as a working hypothesis, not validated (Section 4.5).
standard math SPICE ephemerides provide accurate celestial positions
SPICE is a standard toolkit; this is a reasonable external foundation (Section 3.4).
domain assumption 360-degree HDR environment map approximates Earthshine
The emissive dome is an approximation of secondary illumination; no comparison to radiative transfer is provided (Section 3.5).
domain assumption Rendered annotations are exact ground truth
Pose, keypoints, depth, and normals are derived directly from rendering, but no reprojection or consistency check is reported (Section 4).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A large-scale, physically-based synthetic dataset for satellite pose estimation." pith.science (2026). https://pith.science/paper/IADE3TUC

@misc{pith2026250612782,
  author       = {Pith},
  title        = {Pith review of: A large-scale, physically-based synthetic dataset for satellite pose estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IADE3TUC}},
  note         = {Machine review of arXiv:2506.12782}
}

read the original abstract

The Deep Learning Visual Space Simulation System (DLVS3) introduces a novel synthetic dataset generator and a simulation pipeline specifically designed for training and testing satellite pose estimation solutions. This work introduces the DLVS3-HST-V1 dataset, which focuses on the Hubble Space Telescope (HST) as a complex, articulated target. The dataset is generated using advanced real-time and offline rendering technologies, integrating high-fidelity 3D models, dynamic lighting (including secondary sources like Earth reflection), and physically accurate material properties. The pipeline supports the creation of large-scale, richly annotated image sets with ground-truth 6-DoF pose and keypoint data, semantic segmentation, depth, and normal maps. This enables the training and benchmarking of deep learning-based pose estimation solutions under realistic, diverse, and challenging visual conditions. The paper details the dataset generation process, the simulation architecture, and the integration with deep learning frameworks, and positions DLVS3 as a significant step toward closing the domain gap for autonomous spacecraft operations in proximity and servicing missions.

Figures

Figures reproduced from arXiv: 2506.12782 by the authors.

**Figure 2.** 360-degree HDR environment maps 3.6 Post-processing pipeline To further bridge the gap between synthetic and real-world imagery, the DLVS3 framework incorporates a comprehensive post-processing pipeline developed using OpenCV. This pipeline applies a suite of visual effects and sensor artifacts to the rendered images, ensuring that the synthetic data more closely resembles the outputs of actual spaceborne cameras. T… view at source ↗

**Figure 4.** HST Keypoint positions The primary purpose of these keypoints is to provide a sparse yet informative representation of the Hubble’s 3D pose and spatial extent. By tracking the 2D projections of these 3D keypoints in the rendered images, it becomes possible to estimate the telescope’s orientation and position relative to the camera. The distribution of these 37 keypoints is designed to capture the overall structure a… view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Different layers in the EXR image 4.1 OpenEXR (.exr) format This is a versatile and powerful image format, particularly well-suited for storing high-dynamic-range (HDR) imagery and multiple image elements within a single file. A key advantage of the EXR format, as util…

**Figure 5.** Figure 5: Post-processing on HST dataset images As detailed in the general section on post-processing, the HST dataset includes a set of standard post-processing effects to enhance realism. Specifically, the following effects were applied: panchromatic filtering, vignetting, blo…

**Figure 6.** Figure 6: Full resolution examples from the dataset [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

[1]

Exploring ai-based satellite pose estimation: from novel synthetic dataset to realistic validation

Fabien Gallet, Christophe Marabotto, and Thomas Chambon. Exploring ai-based satellite pose estimation: from novel synthetic dataset to realistic validation. In CVPR 2024 AI4Space Workshop, 2024

work page 2024
[2]

Spacecraft Pose Estimation Based on Unsupervised Domain Adaptation and on a 3D-Guided Loss Combination

José Ignacio Bravo Pérez-Villar et al. Spacecraft pose estimation based on unsupervised domain adaptation and on a 3d-guided loss combination. arXiv preprint arXiv:2212.13415, 2022

work page Pith review arXiv 2022
[3]

Towards bridging the space domain gap for satellite pose estimation using event sensing

Mohsi Jawaid, Ethan Elms, Yasir Latif, and Tat-Jun Chin. Towards bridging the space domain gap for satellite pose estimation using event sensing. arXiv preprint arXiv:2209.11945, 2022

arXiv 2022
[4]

Satellite Pose Estimation Challenge: Dataset, Competition Design and Results

Mate Kisantal, Sumant Sharma, Tae Ha Park, Dario Izzo, Marcus Märtens, and Simone D’Amico. Satellite pose estimation challenge: Dataset, competition design and results. IEEE Transactions on Aerospace and Electronic Systems , 2020. arXiv:1911.02050

work page Pith review arXiv 2020
[5]

Satellite pose estimation with deep landmark regression and nonlinear pose refinement

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 (ICCVW), 2019

work page 2019
[6]

Speed+: Next-generation dataset for spacecraft pose estimation across domain gap

Tae Ha Park, Marcus Märtens, Guillaume 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, 2022

work page 2022
[7]

Adaptive neural-network-based unscented kalman filter for robust pose tracking of noncooperative spacecraft

Tae Ha Park and Simone D’Amico. Adaptive neural-network-based unscented kalman filter for robust pose tracking of noncooperative spacecraft. Journal of Guidance, Control, and Dynamics , 2023. Published Online: 11 Jul 2023

work page 2023
[8]

SPADES: A Realistic Spacecraft Pose Estimation Dataset using Event Sensing

Arunkumar Rathinam, Haytam Qadadri, and Djamila Aouada. Spades: A realistic spacecraft pose estimation dataset using event sensing.arXiv preprint arXiv:2311.05310, 2023

work page Pith review arXiv 2023

Show all 19 references

[9]

Spark: Spacecraft recognition leveraging knowledge of space environment

Mohamed Musallam Adel, Kassem Al Ismaeil, Oyebade Oyedotun, Marcos Damian Perez, Michel Poucet, and Djamila Aouada. Spark: Spacecraft recognition leveraging knowledge of space environment. arXiv preprint arXiv:2104.05978 , 2021

2021 arXiv
[10]

Deep learning for spacecraft pose estimation from photorealistic rendering

Pedro F Proenca and Yang Gao. Deep learning for spacecraft pose estimation from photorealistic rendering. arXiv preprint arXiv:1907.04298, 2019

1907 arXiv
[11]

Sispo: Space imaging simulator for proximity operations

Iaroslav Iakubivskyi and Mihkel Pajusalu. Sispo: Space imaging simulator for proximity operations. PLOS ONE, 17(3):e0263882, 2022

2022
[12]

The surrender software

Pierre Delattre, Vincent Brochard, Nicolas Schmutz, Philippe Roux, Jean-Marc Christille, Walter Riedler, Florian Renk, Fabien Gachet, Jean-Pierre Lebreton, and Giampiero Naletto. The surrender software. In Proceedings of the 8th International Workshop on Satellite Constellatio...

2021
[13]

Sanchez-Gestido

M. Sanchez-Gestido. Pangu version v8: Event-based sensor simulation for space applications in real-time. In ADCSS 2024, European Space Agency , 2024. ESA/TEC-SAG Presentation

2024
[14]

Spin: An open simulator of realistic spacecraft navigation imagery

Javier Montalvo, Juan Ignacio Bravo Pérez-Villar, Álvaro García-Martín, Pablo Carballeira, and Jesús Bescós. Spin: An open simulator of realistic spacecraft navigation imagery. arXiv preprint arXiv:2406.07500, 2024

2024 arXiv
[15]

Houdini: Advanced 3d procedural software for modeling, animation, and rendering

SideFX. Houdini: Advanced 3d procedural software for modeling, animation, and rendering. In ACM SIGGRAPH 2019 Real-Time Live!, 2019. Software presentation

2019
[16]

Unreal engine: Real-time 3d creation tool

Epic Games. Unreal engine: Real-time 3d creation tool. https://www.unrealengine. com/, 2023. Software, version 5. Available at: https://www.unrealengine.com/

2023
[17]

Materialx: An open standard for material and look-development content

MaterialX Development Team. Materialx: An open standard for material and look-development content. https://materialx.org, 2017. Version 1.39.3. Available at: https://materialx.org

2017
[18]

Charles H. Acton. Ancillary data services of nasa’s navigation and ancillary information facility. Planetary and Space Science , 44(1):65–70, 1996

1996
[19]

DLVS3 HST Pose Estimation Dataset

Machine Intelligence Zrt. DLVS3 HST Pose Estimation Dataset. https://mi.services/ dlvs3-hst-dataset-access/ , 2025. Accessed: 2025-06-12. 8

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Exploring ai-based satellite pose estimation: from novel synthetic dataset to realistic validation

Fabien Gallet, Christophe Marabotto, and Thomas Chambon. Exploring ai-based satellite pose estimation: from novel synthetic dataset to realistic validation. In CVPR 2024 AI4Space Workshop, 2024

work page 2024

[2] [2]

Spacecraft Pose Estimation Based on Unsupervised Domain Adaptation and on a 3D-Guided Loss Combination

José Ignacio Bravo Pérez-Villar et al. Spacecraft pose estimation based on unsupervised domain adaptation and on a 3d-guided loss combination. arXiv preprint arXiv:2212.13415, 2022

work page Pith review arXiv 2022

[3] [3]

Towards bridging the space domain gap for satellite pose estimation using event sensing

Mohsi Jawaid, Ethan Elms, Yasir Latif, and Tat-Jun Chin. Towards bridging the space domain gap for satellite pose estimation using event sensing. arXiv preprint arXiv:2209.11945, 2022

arXiv 2022

[4] [4]

Satellite Pose Estimation Challenge: Dataset, Competition Design and Results

Mate Kisantal, Sumant Sharma, Tae Ha Park, Dario Izzo, Marcus Märtens, and Simone D’Amico. Satellite pose estimation challenge: Dataset, competition design and results. IEEE Transactions on Aerospace and Electronic Systems , 2020. arXiv:1911.02050

work page Pith review arXiv 2020

[5] [5]

Satellite pose estimation with deep landmark regression and nonlinear pose refinement

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 (ICCVW), 2019

work page 2019

[6] [6]

Speed+: Next-generation dataset for spacecraft pose estimation across domain gap

Tae Ha Park, Marcus Märtens, Guillaume 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, 2022

work page 2022

[7] [7]

Adaptive neural-network-based unscented kalman filter for robust pose tracking of noncooperative spacecraft

Tae Ha Park and Simone D’Amico. Adaptive neural-network-based unscented kalman filter for robust pose tracking of noncooperative spacecraft. Journal of Guidance, Control, and Dynamics , 2023. Published Online: 11 Jul 2023

work page 2023

[8] [8]

SPADES: A Realistic Spacecraft Pose Estimation Dataset using Event Sensing

Arunkumar Rathinam, Haytam Qadadri, and Djamila Aouada. Spades: A realistic spacecraft pose estimation dataset using event sensing.arXiv preprint arXiv:2311.05310, 2023

work page Pith review arXiv 2023