Joint attitude estimation and 3D neural reconstruction of non-cooperative space objects

Cl\'ement Forray; Dawa Derksen; Florence Genin; Nicolas Delaygue; Pauline Delporte

REVIEW 3 major objections 6 minor 42 references

Joint attitude estimation and 3D neural reconstruction of non-cooperative space objects

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read With known acquisition times, optimizing a single 3D rotation jointly with a neural radiance field reconstructs a uniformly rotating satellite's shape and attitude from monocular images alone, reaching a median pose error of 1.6 degrees.

desk verdict A credible, well-scoped NeRF pipeline for space-object attitude reconstruction, but the key global-vs-indep comparison is confounded by a simultaneous change in training schedule, so the claimed advantage isn't yet established. read the letter →

arxiv 2506.20638 v1 pith:ZIVMAFP3 submitted 2025-06-25 cs.CV

Clément Forray , Pauline Delporte , Nicolas Delaygue , Florence Genin , Dawa Derksen This is my paper · ORCID

classification cs.CV

keywords Neuralradiancefieldsattitudeestimationnon-cooperativespaceobjectssituationalawarenesspose3Dreconstructionuniformrotationmonocularimagesequence

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

This paper aims to show that a neural radiance field (NeRF) can reconstruct the 3D shape of a non-cooperative, uniformly rotating satellite from monochromatic images while simultaneously recovering the satellite's attitude, without any prior pose information, masks, or depth maps. The central move is to exploit the target's uniform rotation: since the image acquisition times are known, every camera pose is expressed by a single 3D axis-angle rotation vector, and that vector is optimized together with the scene network by gradient descent. On a simulated sequence of 99 training images, the global single-rotation mode reaches a median angular pose error of 1.6 degrees, a maximal error of 3 degrees, and a reconstruction precision of 17.2 mm, clearly beating both per-image pose optimization and a structure-from-motion pipeline. A sympathetic reader would care because this addresses the core information gap of space situational awareness — determining what an uncooperative object looks like and how it is tumbling — from the only data that is realistically available.

What carries the argument

The load-bearing device is the global uniform-rotation parameterization: a 3-parameter axis-angle vector $\mathbf{r}$ whose norm is the rotation angle, with camera $i$'s rotation given by the acquisition-time-scaled value $t_i \mathbf{r}$. This reduces the pose search from $6N$ unknowns to 3 and guarantees that consecutive poses differ only slightly, which is exactly what the joint optimization needs. The optimization is carried by an incremental training schedule in which the NeRF first sees eight views, then a new view every 100 steps, combined with a gradient-scaled multi-resolution hash encoding that masks high-frequency gradients early, a logarithmic tone mapping to compress specular radiance, opacity and radiance regularization losses, a distortion loss, and an L1 penalty on the rotation parameters. The single rotation is what allows initialization at zero rotation despite the roughly 15-degree convergence limit of per-image pose refinement.

What would settle it

Run the identical training recipe on a sequence in which the target's angular velocity changes midway (for example, a 30% speed change at frame 50, or a 10-degree axis shift); if the global single-rotation optimization cannot keep pose errors low and the point cloud degrades, then the uniform-rotation assumption is load-bearing. A second falsifier: replace the fixed camera–Sun configuration by a sequence where the Sun angle changes over the pass; if reconstruction quality collapses, the lighting correlation assumption is necessary.

Watch

Extended reading notes

Core claim

The paper claims that, in the specific but realistic scenario of a rigid satellite rotating at constant angular velocity around a fixed axis, observed by a camera fixed relative to the Sun, the attitude and geometry problems collapse into a single optimization. All 99 camera orientations are parameterized as one axis-angle rotation vector, scaled by acquisition time, and a coarse-to-fine gradient-scaled hash encoding keeps early pose gradients smooth while the model learns low frequencies first; new views are then introduced one at a time so each pose starts near its predecessor. Starting from zero rotation and with no prior, this global mode recovers poses with 1.6 degrees median and 3 degrees maximum angular error, produces point-cloud geometry with 17.2 mm precision, and keeps novel-view metrics close to the baseline trained with ground-truth poses. The same model trained with per-image independent rotations reaches only 5 degrees median and 15 degrees maximum, and the structure-from-motion baseline yields 83 degrees median and 174 degrees maximum, which is the paper's evidence that the single-rotation parameterization, not the NeRF itself, is what makes the problem tractable.

Load-bearing premise

The load-bearing premise is that the target rotates at constant angular velocity around a fixed axis for the entire sequence and that the camera and Sun are fixed relative to each other, so the lighting direction is fully correlated with viewing angle; if the real rotation is non-uniform or the lighting geometry changes, the global parameterization is misspecified and the method needs coarse initial poses or a different motion model.

Editorial extensions

If this is right

With known acquisition times, pose estimation for a uniformly rotating object costs only 3 parameters, making joint NeRF training stable enough to start from zero rotation.
The method needs no prior pose estimates, feature matching, masks, or depth maps, so it applies to non-cooperative objects where such auxiliary data is unavailable.
The learned radiance field can be exported as a point cloud or mesh, giving an exploitable 3D model for inspection planning, collision avoidance, or anomaly detection.
Pose errors of 1.6 degrees median keep novel-view synthesis close to the ground-truth-pose baseline, so the same model can be used to predict views not yet seen.

Reading between the lines

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

A natural extension the paper leaves implicit: split a non-uniform tumble into uniform segments and run the global rotation per segment, with segment boundaries detected from image flow; nothing in the method obviously prevents this, but the paper does not test it.
The fixed-camera-behind-Sun configuration is an idealization. In a real pass the Sun angle changes with orbit geometry, so a testable variant would jointly estimate the Sun direction as one more parameter and check whether pose accuracy survives.
The reported sensitivity to hyperparameters (regularization weights, hash-encoding schedule, pose learning rate) suggests the method is not yet turnkey; a robust schedule or automatic tuning would be the difference between a demonstration and an operational tool.
Because the single-rotation parameterization encodes the motion model itself, the 3-degree maximum pose error also gives a running estimate of the target's spin axis and rate, which could feed a Kalman filter for attitude prediction—an operational benefit the paper does not discuss.

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

Worth a read if you work on NeRF-based reconstruction of uncooperative space objects. The paper combines known ingredients in a sensible way: log tone mapping, opacity/radiance regularization, gradient-scaled multi-resolution hash encoding, an incremental view schedule, and a global 3-parameter uniform-rotation parameterization using known acquisition times. That last piece is the most interesting practical idea: instead of per-image poses, you fit one axis-angle vector and derive all cameras from it. On their simulated Galileo sequence the global mode clearly beats per-image pose optimization and SfM in both pose error and 3D precision, while roughly matching baseline radiometric quality.

The soft spot is the central comparison. Global differs from Indep in two ways at once: the shared rotation parameterization and the two-stage incremental schedule with gradient-scaled hash encoding. Indep starts with poses already within ~8 degrees and apparently no incremental view introduction. So the global-mode advantage could come from the schedule or the favorable initialization rather than from the uniform-rotation parameterization itself. The abstract and conclusion attribute the result to "optimizing a uniform rotation," but that attribution is underdetermined. The dataset is also generated under exactly that assumption (uniform rotation, camera fixed relative to the Sun), so the global model is given a structural advantage that the paper discloses but never stress-tests against non-uniform motion.

The paper is honest about these limits in the Limitations section, which earns some credit. Still, evaluation is a single simulated sequence, there is no comparison to CF-NeRF or LU-NeRF, and no code or data are released. All of these are addressable. An ablation that runs global and indep under the same incremental schedule, plus an experiment on a mildly non-uniform rotation, would substantially firm up the claim.

Who is this for? Practitioners in space situational awareness who want a baseline for monocular, sun-fixed inspection scenarios, and NeRF researchers curious about explicit motion models. The problem is real and the basic approach is plausible. I would not desk-reject it; the confound is a major but fixable issue, not a reason to reject the underlying idea. Send it to a serious referee, ideally with expectations of major revision.

Referee Report

3 major / 6 minor

Summary. The paper proposes a NeRF-based method for joint 3D reconstruction and attitude estimation of non-cooperative space objects from a sequence of monocular, monochromatic images. The method assumes the target rotates uniformly about a fixed axis, and with known acquisition times it parameterizes all camera poses as a single 3-parameter rotation that is optimized jointly with the NeRF. It combines an incremental view-introduction schedule, a gradient-scaled hash encoding, logarithmic tone mapping, and several regularization losses. Experiments on one simulated sequence of a Galileo satellite (99 training images, 10 validation images) report that this global mode achieves median and maximal angular pose errors of 1.6° and 3°, respectively, and a 3D reconstruction precision of 17.2 mm, outperforming per-image pose optimization (indep) and a SfM pipeline, with radiometric quality close to the true-pose baseline.

Significance. If the results hold, the paper offers an interesting demonstration that a physics-informed motion prior can stabilize joint NeRF optimization and pose estimation in a challenging space-imaging setting. The work is honest about its assumptions and limitations, and it provides quantitative pose and geometry metrics rather than qualitative results alone. The strongest feature is the explicit use of the uniform-rotation prior, which is a natural and well-motivated inductive bias for tumbling or spinning space objects. However, the claimed advantage over per-image pose optimization is currently underdetermined by the experiments because the global and indep conditions differ in more than one variable, and the evaluation is limited to a single simulated sequence generated under exactly the motion model assumed by the method. The central claim is defensible, but it needs additional ablations and robustness experiments to be fully established.

major comments (3)

[§7.1, Tables 2–3] The global and indep conditions differ in two simultaneous respects: (i) global estimates a single 3-parameter rotation for all cameras, while indep optimizes one rotation per image; and (ii) global uses a two-stage incremental schedule (first 8 views, then one new view every 100 steps) with a zero-initialized rotation, while indep initializes all poses near the truth (Gaussian noise with mean 8° and standard deviation 2°) and does not appear to use an incremental view schedule. The reported advantage of global over indep (median angular error 1.6° vs 5°, maximal 3° vs 15°, precision 17.2 mm vs 26.2 mm) could therefore be due to the training schedule, the initialization, or the parameterization, rather than to the uniform-rotation model itself. Since the abstract and conclusion attribute the improvement to the uniform-rotation optimization, please provide an ablation that separates these factors, for example indep with the same incremental schedule and zero initialization, or global with per-image rotations.
[§3 and §5.2] The dataset is generated under precisely the motion model assumed by the global parameterization: a uniform rotation about a fixed axis, with the camera fixed relative to the Sun so that lighting is perfectly correlated with viewing angle. The conclusion in §8 states that 'when this hypothesis is not verified, our experiments demonstrate that the model could work with coarse camera poses,' but no experiment with non-uniform motion (e.g., angular acceleration, axis drift, or a changing angular velocity) is reported. Because the central contribution is attitude estimation for non-cooperative objects, the method's robustness to deviations from the assumed motion is a load-bearing concern. Please add a stress test with perturbed motion, or temper the conclusion to reflect that only uniform-rotation cases were evaluated.
[Tables 2–3] The tables report means and standard deviations, but it is unclear whether these statistics are computed over validation images, over point-cloud points, or over multiple independent training runs with different random seeds. With one simulated sequence and apparently one run per condition, the statistical significance of the global-vs-indep differences is not established. Please clarify the source of the variance and, ideally, provide repeated-seed runs or at least state the number of runs on which each mean and standard deviation are based.

minor comments (6)

[Abstract, Table 1, §7.2] There are several typographical errors: 'primarly' in the abstract, 'PNSR' in the Table 1 header (should be 'PSNR'), and 'shows shows' in Section 7.2.
[§5.2] The text says 'the proposed scaled hash encoding' but the method is an adaptation of the gradient-scaled encoding from reference [14]. Please state explicitly what is new beyond that reference to avoid overclaiming novelty.
[§5.4] The L1 regularization on camera rotation parameters is mentioned but its weight is not specified, nor are the weights for the opacity and radiance losses. Please list these hyperparameters, as the paper's own Limitations section identifies them as critical to the results.
[§5.3, Eq. (5)] The tone-mapping normalization constant M is described only as 'greater than the maximum radiance value in the dataset.' Please state how M is chosen in practice, since the results may depend on this choice.
[§7.4] For the SfM baseline, the paper reports an angular error of 174° maximal and 83° median but does not specify which SfM implementation was used, whether it was given the same 99 images, or whether feature matching failed. Please provide these details so the comparison is reproducible.
[§3] The sequence contains 109 images covering 'slightly more than a full rotation.' Please state the approximate angular step between consecutive views, which is relevant to the assumed uniform-rotation parameterization and to the incremental schedule.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: pose estimates are fit to external ground truth under a disclosed motion model; the global-vs-indep confound is an experimental-design issue, not a logical reduction.

full rationale

The paper's central validation is not circular. Global mode parameterizes all camera poses as a single 3D rotation using known acquisition times (Sec. 5.2), and the simulated dataset was generated under the same uniform-rotation assumption (Sec. 3), so the model family matches the data generator. But the rotation parameters themselves are unknown and learned from image radiance; angular errors in Table 3 are measured against independent pose metadata, and reconstruction precision in Table 2 is measured against the reference mesh. No fitted parameter is renamed as a prediction: the reported pose errors are training-set fit residuals, not forecasts, and the paper does not claim otherwise. The only related concern is that the 'Global' vs 'Indep' comparison changes both the pose parameterization and the training schedule (Sec. 7.1), so the stated advantage is underdetermined; this is a benchmarking/confounding issue, not circularity. Citations to the authors' prior work [19] are contextual and not load-bearing for the pose-estimation claim. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central result depends on a small number of optimized pose parameters, hand-tuned regularization weights, and strong domain assumptions about uniform rotation and fixed lighting geometry. The appearance embedding is a learned latent with no external evidence. No new physical entities are introduced.

free parameters (4)

Global uniform rotation vector (axis-angle, 3 parameters) = median pose error 1.6 deg, max 3 deg
Optimized jointly with the NeRF field to align all camera poses; in global mode this single vector encodes the whole sequence motion.
Per-image pose rotations (indep mode) = median 5 deg, max 15 deg angular error
Independent 3D rotations per image, initialized with Gaussian noise (mean 8 deg, std 2 deg), used as comparison.
Regularization weights and schedule hyperparameters = not specified numerically
Opacity, radiance, distortion and rotation L1 losses plus schedule length and learning rates are hand-tuned; authors say results depend critically on them in Section 9.
Per-image appearance embedding = learned per image index, no value given
A learned code added to the radiance model to absorb view inconsistencies; described in Section 5.1.

assumptions (5)

domain assumption The target object rotates uniformly about a fixed axis and image acquisition times are known
Section 5.2 parameterizes all cameras by one 3-parameter rotation; Section 7.1 global experiment relies on it. A non-uniform motion breaks the parameterization.
domain assumption Camera intrinsics and observer position/orientation are known
Section 3 states these are known during the sequence; only the target attitude is unknown.
domain assumption The Sun is fixed behind the camera and there is no diffuse lighting
Section 3 and Section 5.2 justify modeling radiance as a function of viewing direction only; shadows and specularities correlate perfectly with pose.
domain assumption Initial poses must be within roughly 15 degrees of the truth
Section 5.2 states the method works only if poses initialize not too far from true positions; the incremental schedule exists to satisfy this.
standard math NeRF volumetric rendering equations (Eqs. 1-4) are valid scene models
The density and radiance field rendering from Mildenhall et al. is the foundation of the loss and point-cloud extraction.

invented entities (1)

Learned per-image appearance embedding
purpose: Absorbs minor radiometric inconsistencies between views to stabilize radiance prediction
Not a physical entity; a standard per-image latent code. It has no observable handle outside the paper's own training data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Joint attitude estimation and 3D neural reconstruction of non-cooperative space objects." pith.science (2026). https://pith.science/paper/ZIVMAFP3

@misc{pith2026250620638,
  author       = {Pith},
  title        = {Pith review of: Joint attitude estimation and 3D neural reconstruction of non-cooperative space objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZIVMAFP3}},
  note         = {Machine review of arXiv:2506.20638}
}

read the original abstract

Obtaining a better knowledge of the current state and behavior of objects orbiting Earth has proven to be essential for a range of applications such as active debris removal, in-orbit maintenance, or anomaly detection. 3D models represent a valuable source of information in the field of Space Situational Awareness (SSA). In this work, we leveraged Neural Radiance Fields (NeRF) to perform 3D reconstruction of non-cooperative space objects from simulated images. This scenario is challenging for NeRF models due to unusual camera characteristics and environmental conditions : mono-chromatic images, unknown object orientation, limited viewing angles, absence of diffuse lighting etc. In this work we focus primarly on the joint optimization of camera poses alongside the NeRF. Our experimental results show that the most accurate 3D reconstruction is achieved when training with successive images one-by-one. We estimate camera poses by optimizing an uniform rotation and use regularization to prevent successive poses from being too far apart.

Figures

Figures reproduced from arXiv: 2506.20638 by the authors.

**Figure 2.** NeRF model representation This representation allows to generate novel views of the scene from any arbitrary angle through volumetric rendering. The value of a pixel in the produced image corresponds to the sum of the radiances of the points along the line of sight weighted by their visibility as shown in Eq. (3). The visibility of a given point being determined by its own density and that of the points separating… view at source ↗

**Figure 3.** Architecture of the model (hash-encoding source [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Precision point cloud (in meters) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Recall point cloud (in meters) 7.4. Pose estimation We compare the estimated camera poses to the poses metadata of our dataset [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Validation view (reference image compared to rendered radiance, accumulation, depth map and surface normal directions) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages

[1]

3d point cloud and mesh processing software open source project. 2024. 3, 5

work page 2024
[2]

Leveraging neural radiance fields for pose estimation of an unknown space object during proximity operations,

Christophe De Vleeschouwer Antoine Legrand, Renaud De- try. Leveraging neural radiance fields for pose estimation of an unknown space object during proximity operations,

work page
[3]

Mip-nerf 360: Unbounded anti-aliased neural radiance fields

Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 5

2022
[4]

NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior

Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural ra- diance field with no pose prior, 2023. Published in arXiv, arXiv:2212.07388. 2

work page Pith review arXiv 2023
[5]

3d reconstruction method of space tar- get on optical images with wide baseline via neural radiance field

Fan Bu, Canyu Wang, Xiaoyuan Ren, Dou Sun, Zhan Wang, and Zhuang Wang. 3d reconstruction method of space tar- get on optical images with wide baseline via neural radiance field. Journal of Physics: Conference Series , 2347(1):99– 106, 2022. 2

work page 2022
[6]

White, and Todd Steffen

Basilio Caruso, Trupti Mahendrakar, Van Minh Nguyen, Ryan T. White, and Todd Steffen. 3d reconstruction of non-cooperative resident space objects using instant ngp- accelerated nerf and d-nerf, 2023. Published in arXiv, preprint arXiv:2301.09060. 2

arXiv 2023
[7]

Marrying nerf with feature matching for one-step pose estimation, 2024

Ronghan Chen, Yang Cong, and Yu Ren. Marrying nerf with feature matching for one-step pose estimation, 2024. Pub- lished in 2024 IEEE International Conference on Robotics and Automation. 2

work page 2024
[8]

Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,

Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,

Show all 42 references

[9]

3d spacecraft recon- struction from on-orbit imagery

Daniel Crispell and Scott Richardson. 3d spacecraft recon- struction from on-orbit imagery. In 3rd Space Imaging Work- shop, 2022. 2

2022
[10]

Launch collision avoidance sensitivity anal- ysis

Gabriele Falco, Federico Toso, Andrew Ratcliffe, and Noelia S´anchez-Ortiz. Launch collision avoidance sensitivity anal- ysis. In Proc. 8th European Conference on Space Debris (virtual), Darmstadt, Germany, pages 20–23, 2021. 1

2021
[11]

Space surveillance and track- ing in esa’s ssa programme

Tim Flohrer and Holger Krag. Space surveillance and track- ing in esa’s ssa programme. In 7th European conference on space Debris, 2017. 1

2017
[12]

Pose estimation and neural implicit re- construction towards non-cooperative spacecraft without of- fline prior information

Bing Han, Chenxi Wang, Xinyu Zhang, Zhibin Zhao, Zhi Zhai, and Jinxin Liu. Pose estimation and neural implicit re- construction towards non-cooperative spacecraft without of- fline prior information. IEEE Transactions on Aerospace and Electronic Systems, pages 1–21, 2024. 2

2024
[13]

Multiple view ge- ometry in computer vision

Richard Hartley and Andrew Zisserman. Multiple view ge- ometry in computer vision . Cambridge university press,
[14]

Kim, and Jin-Hwa Kim

Hwan Heo, Taekyung Kim, Jiyoung Lee, Jaewon Lee, Soohyun Kim, Hyunwoo J. Kim, and Jin-Hwa Kim. Robust camera pose refinement for multi-resolution hash encoding,
[15]

On optimal observation orbits for learning gaussian splatting-based 3d models of unknown resident space objects

Arianna Issitt, Trupti Mahendrakar, Amy Alvarez, Ryan T White, and Alex Sizemore. On optimal observation orbits for learning gaussian splatting-based 3d models of unknown resident space objects. In AIAA SCITECH 2025 Forum, page 1780, 2025. 2

2025
[16]

Barf: Bundle-adjusting neural radiance fields,

Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Si- mon Lucey. Barf: Bundle-adjusting neural radiance fields,
[17]

Published in International Conference on Machine Learning. 4
[18]

Fast digital orthophoto generation: A comparative study of ex- plicit and implicit methods

Jianlin Lv, Guang Jiang, Wei Ding, and Zhihao Zhao. Fast digital orthophoto generation: A comparative study of ex- plicit and implicit methods. Remote Sensing, 16, 2024. 4

2024
[19]

Vision-based neural scene representations for space- craft

Anne Mergy, Gurvan Lecuyer, Dawa Derksen, and Dario Izzo. Vision-based neural scene representations for space- craft. In CVPRW, 2021. 2

2021
[20]

Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65:99–106, 2021. 2, 3

2021
[21]

An active debris removal parametric study for leo environment remediation

J-C Liou. An active debris removal parametric study for leo environment remediation. Advances in space research , 47 (11):1865–1876, 2011. 1

2011
[22]

Space-based space surveillance oper- ational and demonstration missions

D Escorial Olmos, Fernando E Alem ´an Roda, Kevin Middle- ton, and Joris Naudet. Space-based space surveillance oper- ational and demonstration missions. In Proceedings of the 6th European Conference on Space Debris, pages 2013–04,

2013
[23]

Guidelines for active removal of non-functional targets de- signed to assist rendezvous and capture

Giovanni B Palmerini, Marco Sabatini, and Paolo Gasbarri. Guidelines for active removal of non-functional targets de- signed to assist rendezvous and capture. In 2016 IEEE Aerospace Conference, pages 1–13. IEEE, 2016. 1

2016
[24]

Speed+: Next-generation dataset for spacecraft pose estimation across domain gap, 2022

Tae Ha Park, Marcus M ¨artens, Gurvan Lecuyer, Dario Izzo, and Simone D’Amico. Speed+: Next-generation dataset for spacecraft pose estimation across domain gap, 2022. Pub- lished in 2022 IEEE Aerospace Conference (AERO). 2

2022
[25]

Instant neural graphics primitives with a mul- tiresolution hash encoding

Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 41(102):1–15, 2022. 3

2022
[26]

Spacecraft collision avoidance technology

Zhang Rongzhi and Yang Kaizhong. Spacecraft collision avoidance technology. Academic Press, 2020. 1

2020
[27]

Graf: Generative radiance fields for 3d-aware im- age synthesis

Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware im- age synthesis. Advances in Neural Information Processing Systems, 33:20154–20166, 2020. 2

2020
[28]

Photo tourism: exploring photo collections in 3d

Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835–846. 2006. 2

2006
[29]

Ct-nerf: Incremental optimiz- ing neural radiance field and poses with complex trajectory,

Yunlong Ran, Yanxu Li, Qi Ye, Yuchi Huo, Zechun Bai, Ji- ahao Sun, and Jiming Chen. Ct-nerf: Incremental optimiz- ing neural radiance field and poses with complex trajectory,
[30]

Published in arXiv, arXiv:2404.13896. 2

arXiv
[31]

On-orbit servicing: A brief survey

Andrew Tatsch, Norman Fitz-Coy, and Svetlana Gladun. On-orbit servicing: A brief survey. In Proceedings of the IEEE International Workshop on Safety, Security, and Res- cue Robotics (SSRR’06) , pages 276–281. Inst. of Electrical and Electronics Engineers New York, 2006. 1

2006
[32]

The viewing-from- above bias and the silhouette illusion

Nikolaus F Troje and Matthew McAdam. The viewing-from- above bias and the silhouette illusion. i-Perception, 1(3): 143–148, 2010. 2

2010
[33]

Cf-nerf: Camera parameter free neural radiance fields with incremental learning, 2023

Qingsong Yan, Qiang Wang, Kaiyong Zhao, Jie Chen, Bo Li, Xiaowen Chu, and Fei Deng. Cf-nerf: Camera parameter free neural radiance fields with incremental learning, 2023. Published in the 38th AAAI Conference on Artificial Intelli- gence (AAAI24). 2

2023
[34]

Model- ing the world from internet photo collections

Noah Snavely, Steven M Seitz, and Richard Szeliski. Model- ing the world from internet photo collections. International journal of computer vision, 80:189–210, 2008. 2

2008
[35]

Nerfstudio: A modular framework for neural radiance field development

Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristof- fersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SI...

2023
[36]

Overview on space-based optical orbit determination method employed for space situational awareness: From theory to application

Zhe Zhang, Gaopeng Zhang, Jianzhong Cao, Cheng Li, Weining Chen, Xin Ning, and Zheng Wang. Overview on space-based optical orbit determination method employed for space situational awareness: From theory to application. In Photonics, page 610. MDPI, 2024. 1
[37]

Deep review and analysis of recent nerfs

Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12(1), 2023. 2

2023
[39]

Neural radiance field- based visual rendering: a comprehensive review

Mingyuan Yao, Yukang Huo, Yang Ran, Qingbin Tian, Ruifeng Wang, and Haihua Wang. Neural radiance field- based visual rendering: a comprehensive review. arXiv preprint arXiv:2404.00714, 2024. 2

2024 arXiv
[40]

3d re- construction of space objects from multi-views by a visible sensor

Haopeng Zhang, Quanmao Wei, and Zhiguo Jiang. 3d re- construction of space objects from multi-views by a visible sensor. Sensors, 17:1689, 2017. 2

2017
[2021]

Published in 2021 IEEE/CVF International Confer- ence on Computer Vision (ICCV). 4

2021
[2023]

Published in 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV). 2

2023
[2024]

Published in IEEE International Conference on Space Robotics 2024. 2

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

3d point cloud and mesh processing software open source project. 2024. 3, 5

work page 2024

[2] [2]

Leveraging neural radiance fields for pose estimation of an unknown space object during proximity operations,

Christophe De Vleeschouwer Antoine Legrand, Renaud De- try. Leveraging neural radiance fields for pose estimation of an unknown space object during proximity operations,

work page

[3] [3]

Mip-nerf 360: Unbounded anti-aliased neural radiance fields

Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 5

2022

[4] [4]

NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior

Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural ra- diance field with no pose prior, 2023. Published in arXiv, arXiv:2212.07388. 2

work page Pith review arXiv 2023

[5] [5]

3d reconstruction method of space tar- get on optical images with wide baseline via neural radiance field

Fan Bu, Canyu Wang, Xiaoyuan Ren, Dou Sun, Zhan Wang, and Zhuang Wang. 3d reconstruction method of space tar- get on optical images with wide baseline via neural radiance field. Journal of Physics: Conference Series , 2347(1):99– 106, 2022. 2

work page 2022

[6] [6]

White, and Todd Steffen

Basilio Caruso, Trupti Mahendrakar, Van Minh Nguyen, Ryan T. White, and Todd Steffen. 3d reconstruction of non-cooperative resident space objects using instant ngp- accelerated nerf and d-nerf, 2023. Published in arXiv, preprint arXiv:2301.09060. 2

arXiv 2023

[7] [7]

Marrying nerf with feature matching for one-step pose estimation, 2024

Ronghan Chen, Yang Cong, and Yu Ren. Marrying nerf with feature matching for one-step pose estimation, 2024. Pub- lished in 2024 IEEE International Conference on Robotics and Automation. 2

work page 2024

[8] [8]

Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,

Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs,