Pith. sign in

REVIEW 3 major objections 4 minor 44 references

Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures

T0 review · 3 major / 4 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 This paper argues that an optimized radially symmetric background disc can force camera pose estimators to collapse their predicted orientations toward a single viewpoint.

desk verdict The kaleidoscopic background attack is empirically strong and novel, but the loss's orientation-consistency derivation has a coordinate-frame error that needs fixing before the mechanism claim holds. read the letter →

arxiv 2507.10265 v1 pith:B3WGZEDW submitted 2025-07-14 cs.CV

Xinlong Ding , Hongwei Yu , Jiawei Li , Feifan Li , Yu Shang , Bochao Zou , Huimin Ma , Jiansheng Chen This is my paper · ORCID
classification cs.CV
keywords adversarialattackcameraposeestimationradialsymmetrykaleidoscopicbackgroundphysical-worldpointmapprojectedorientationconsistencysparse-view3Dvision
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 tries to show that a physical background texture alone, arranged as a radially symmetric kaleidoscopic disc, can break modern learning-based camera pose estimators in object-centric scenes. It constructs the disc by repeating one texture segment N times around a circle, so every viewpoint sees nearly the same background; natural texture segments already degrade pose estimates, and an optimized version is dramatically stronger. The optimization maximizes a projected orientation consistency loss that aligns the direction in which pointmap coordinates increase across the disc, a surrogate for aligning the cameras' estimated orientation vectors. If the claim holds, sparse-view pose estimation systems are vulnerable to a cheap physical-world attack, and robustness work on such systems must take scene background geometry seriously.

What carries the argument

The load-bearing object is the kaleidoscopic disc: N identical segment images warped by perspective projections into wedges and summed into a 2ρ×2ρ background (Eqs. 5–6), so the pattern is invariant under rotation by 2π/N. The attack objective is the projected orientation consistency loss Lpoc (Eq. 16), computed from DUSt3R's pointmaps: for each channel, three bisecting lines divide the disc, the average coordinate offset across each division defines a flow direction τ̄i, and the loss maximizes the cosine similarity of these flow directions between two views. The link to pose is the identity that the in-plane gradient of the pointmap channel is the i-th row of the rotation matrix, and its projection onto the background plane is approximated by the measured flow direction; maximizing Lpoc therefore approximately aligns the camera orientation vectors projected onto the disc plane (Eqs. 17–22).

What would settle it

Render a set of random backgrounds, compute the optimized disc for a subset, then test a DUSt3R variant fine-tuned on images whose backgrounds are masked out before training; if pose accuracy under KBAopt stays near the natural-background level, the attack is exploiting a background shortcut rather than an inescapable geometric failure. A cheaper check is to compute Lpoc and the true relative rotation error over many random discs: if maximizing Lpoc does not reliably increase RRA, the surrogate identity in Eq. 21 is broken.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-fold radial symmetry in the imaged background is itself an adversarial structure: a disc built from N identical perspective-warped segments stays visually similar across cameras pointed at the same object, and pose models lean on that similarity. In physical experiments with printed discs around real objects, the unoptimized kaleidoscopic background (KBAnat) lowers pose accuracy on six models, and the optimized KBAopt collapses DUSt3R's relative rotation accuracy RRA@15 from 1.00 with a natural desktop background to 0.00 while raising relative rotation similarity RRS from 0.62 to 0.94. The same printed disc transfers to black-box models MASt3R, RayDiffusion, RayRegression, RelPose++, and PoseDiffusion, with KBAopt yielding lower and more consistent metrics than both the natural and the unoptimized backgrounds.

Load-bearing premise

The attack stands on the assumption that the direction in which the model's per-pixel 3D coordinates change across the background disc faithfully represents the camera's projected orientation vectors, so that aligning those directions across views really does collapse the estimated poses; if that surrogate is wrong for a given model, the optimized texture may only exploit the quirks of one model's coordinate statistics rather than a general geometric weakness.

Editorial extensions

If this is right

  • A background printed on a flat disc is sufficient to degrade pose estimation in the physical world; no modification of the object itself is needed.
  • Optimizing only for orientation consistency also collapses predicted camera positions, so downstream reconstruction from the attacked views is severely disrupted.
  • The attack works with 3, 5, or 10 input views and across varying camera distances and pitch angles, not only at the exact optimization configuration.
  • Black-box transfer holds across models with different output representations, including pointmap-based and ray-based pose estimators.
  • Even unoptimized radially symmetric versions of natural desktop textures already reduce pose accuracy, indicating that radial symmetry alone carries attack power.

Reading between the lines

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

  • If the failure is caused by the model's reliance on background texture as a pose cue, then training pose estimators with random radial-symmetry backgrounds or foreground-masked crops should substantially reduce the attack; this is a direct test the paper does not run.
  • The same disc geometry might also confuse classical feature-based matching, since radial symmetry makes local patches near the disc center ambiguous under rotation; whether SfM pipelines are affected is left open.
  • Because optimization enforces consistency of the projected orientation only on the disc plane, the attack should weaken when the disc covers less of the image, when the object occludes the center, or when the camera looks steeply down; measuring KBAopt under those conditions would delimit the threat.
  • KBAopt could be repurposed as a physical-world stress test for pose estimation systems deployed in robotics and augmented reality, where an operator could place such a disc in a scene to audit whether localization is relying on background shortcuts.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes attacking sparse-view camera pose estimators in object-centric scenes by placing a radially symmetric 'kaleidoscopic' disc in the background. The disc is built from N repeated segments, and an optimized version (KBAopt) is obtained by maximizing a projected orientation consistency loss Lpoc (Eq. 16) against DUSt3R as a white-box surrogate. The authors report digital and physical experiments across DUSt3R, MASt3R, RayDiffusion, RayRegression, RelPose++, and PoseDiffusion, showing large drops in RRA/RTA/mAA with KBAopt and improved orientation consistency as measured by their RRS metric. Ablations compare natural textures, plain radial symmetry, and alternative optimization losses.

Significance. If the central claim holds, the paper identifies a realistic and previously underexplored vulnerability: pose estimation models that rely on background cues can be disrupted by a physically printable, radially symmetric texture that looks similar from multiple views but is optimized to collapse predicted orientations. The empirical stack is a strength: the authors include both digital rendering and physical capture, white-box and black-box transfer, six evaluated models, standard pose metrics, and ablations. The main weakness is the theoretical interpretation of Lpoc in Eqs. 21-22, which is asserted rather than established and appears internally inconsistent with the stated DUSt3R output convention. The work is potentially publishable, but the mechanism behind the loss needs to be re-derived or explicitly reframed as an unproven heuristic, with appropriate empirical validation.

major comments (3)
  1. [Section 2.2, Eqs. (18)-(22)] The derivation of Lpoc as a surrogate for projected orientation similarity is not valid under the manuscript's own description of DUSt3R. The text states immediately after Eq. (10) that 'DUSt3R ... simultaneously regresses the pointmaps Oa and Ob in the camera coordinate system of a.' The interpretation in Eqs. (18)-(21) instead assumes that each pointmap is expressed in its own camera coordinate system, so that the Jacobian of the coordinate map is the view's own rotation matrix R_b and hence ∇Φ_i = r_i^b. For view b under the stated convention, the coordinate function is the camera-a coordinate map evaluated at points imaged through view b, not the camera-b coordinate map. Therefore the image-space flow τ̄_i^b is not a projection of r_i^b, and Eq. (22) does not follow. This is load-bearing because Lpoc is presented as the theoretical reason why the attack induces orientation collapse. The authors should either correct the coordinate-frame derivation, or explicitly state that Lpoc is an empirical surrogate that is not currently derived, and support that weaker claim with evidence that Lpoc correlates with actual pose error.
  2. [Section 2.2, Eq. (21)] Even setting aside the coordinate-frame issue, Eq. (21) relies on several unverified approximations: orthogonal projection, a planar background, and finite differences over three 30-degree bisection lines used as a proxy for the plane gradient. These assumptions are not interrogated in the experiments. Since Lpoc is claimed to be 'a fair approximation' of the projected orientation cosine similarity, the paper should provide a direct validation: for example, compute the correlation between Lpoc (or its per-component values) and the true relative rotation error or orientation consistency on a held-out set of rendered scenes, both for random backgrounds and along an optimization trajectory. The current paper only shows that optimizing Lpoc improves attack metrics, which is consistent with Lpoc being a heuristic that happens to work rather than with the specific orientation-similarity interpretation.
  3. [Section 3.2 and Table 1 caption] The physical-world results are presented as the main demonstration of the abstract's claim, but the caption of Table 1 is ambiguous: 'Each cell contains two values: the larger value represents the mean ... while the smaller value indicates the standard deviation.' This makes it difficult to assess whether the reported differences between Nature, KBAnat, and KBAopt are stable across the 24 objects. The standard deviation is reported only across object categories, not across capture sessions or random seeds. Please clarify the experimental design and report a measure of variability across repeated capture/optimization runs, at least for the flagship DUSt3R rows.
minor comments (4)
  1. [Algorithm 1, line 5] The rendering call uses 'I0_s' as the texture input instead of 'I_t_s'; this appears to be a typo, since the constructed disc Id already contains the current segment image I_t_s. Please correct it.
  2. [Section 2.2, Eq. (16) and Eq. (21)] The definition of τ̄_i in Eq. (15) omits the normalization by distance ∥s1-s2∥ that appears in Eq. (21). If the two are intended to be the same quantity, the missing factor should be explained; if they are intentionally different, the text should avoid calling τ̄_i a gradient estimate without qualification.
  3. [Section 3, evaluation metrics] The custom RRS metric measures only consistency of predicted relative rotations and not accuracy relative to ground truth. This is fine as an auxiliary diagnostic, but the paper should state this limitation explicitly when interpreting 'orientation similarity close to 0.9' as a successful attack.
  4. [References and notation] The notation 'RDF (right-down-forward) convention' is credited to the OpenCV reference [4], but the rotation matrix convention in Eq. (7) is not explicitly connected to that page. A brief statement of the row/column convention would clarify the derivation in Eqs. (18)-(21).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the attack's optimization objective and its evaluation metrics are external to one another.

full rationale

The central claim is that kaleidoscopic backgrounds, especially KBAopt, disrupt camera pose estimation. KBAopt is optimized against the surrogate loss Lpoc (Eq. 16), which is computed from DUSt3R pointmaps, and is then evaluated with standard pose metrics (RRA, RTA, mAA) and a proposed RRS metric on held-out objects, scenes, and black-box models. The evaluation metrics are not the optimization objective: RRA/RTA/mAA measure angular error against ground truth, and RRS measures consistency of predicted relative rotations. None of these is a fitted parameter renamed as a prediction, and no constant is fitted to the test data. The radial-symmetry construction is a design prior, not a result derived from the data. The paper does cite the authors' own prior work in the related-work section (e.g., refs. [11,12,25,42]), but these citations are not load-bearing for the attack mechanism or the evaluation. The mathematical validity of the surrogate justification in Eqs. 21-22 may be a correctness concern—if the coordinate-frame argument is wrong, Lpoc could be optimizing a DUSt3R-specific pointmap statistic rather than true orientation consistency—but that is an issue of whether an approximation is valid, not circularity, because the empirical evaluation remains external to the optimized objective. Therefore, no specific circular step can be exhibited under the required standard.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No physically invented entities. The free parameters are design hyperparameters of the attack, not fitted to external benchmarks. The key axioms are the DUSt3R pointmap-to-orientation approximation and the assumption that pose estimators use background information; both are stated in the paper. The perspective homography construction relies on standard OpenCV functionality.

free parameters (5)
  • number of radial segments N = 12
    Chosen by hand; affects the symmetry angle theta = 30 degrees and the construction in Eq. 1.
  • line count and angle for flow estimation = 3 lines at 30 degrees spacing
    Chosen in Eq. 15 to handle occlusion; not optimized.
  • FGSM step size alpha = 1/255
    Standard adversarial attack step, chosen by hand.
  • CMYK color-space clipping frequency Tc = not stated in main text
    Added for physical-world printability; frequency chosen by hand.
  • disc radius rho = 1 m digital; 15/20 cm physical
    Scene-specific design choice.
assumptions (4)
  • domain assumption DUSt3R's pointmap channels encode the scene coordinates in the first camera's coordinate system, so gradients of the pointmap with respect to pixel coordinates reflect the camera orientation rows.
    Used in Section 2.2 to define flow directions tau_i and connect them to R rows.
  • domain assumption The gradient of the camera coordinate mapping in the disc plane can be approximated by finite differences of segment centroid values (Eq. 20-21), with orthogonal projection and a planar disc.
    This approximation justifies Lpoc; the paper calls it a 'fair approximation' and a 'relaxed version'.
  • domain assumption Sparse-view object-centric pose estimation relies on background texture information, so background changes can influence the predicted pose.
    Motivates the entire attack; stated in Section 1.
  • standard math Perspective projection homographies computed via OpenCV getPerspectiveTransform correctly map the source segment to the destination quadrilateral for each segment.
    Construction in Section 2.1 relies on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures." pith.science (2026). https://pith.science/paper/B3WGZEDW

@misc{pith2026250710265,
  author       = {Pith},
  title        = {Pith review of: Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B3WGZEDW}},
  note         = {Machine review of arXiv:2507.10265}
}
read the original abstract

Camera pose estimation is a fundamental computer vision task that is essential for applications like visual localization and multi-view stereo reconstruction. In the object-centric scenarios with sparse inputs, the accuracy of pose estimation can be significantly influenced by background textures that occupy major portions of the images across different viewpoints. In light of this, we introduce the Kaleidoscopic Background Attack (KBA), which uses identical segments to form discs with multi-fold radial symmetry. These discs maintain high similarity across different viewpoints, enabling effective attacks on pose estimation models even with natural texture segments. Additionally, a projected orientation consistency loss is proposed to optimize the kaleidoscopic segments, leading to significant enhancement in the attack effectiveness. Experimental results show that optimized adversarial kaleidoscopic backgrounds can effectively attack various camera pose estimation models.

Figures

Figures reproduced from arXiv: 2507.10265 by the authors.

Figure 1
Figure 1. Impact of natural and kaleidoscopic backgrounds on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Construction of the kaleidoscopic background disc. (b) Estimation of coordinate flow direction. (c) Calculation of average [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) The setup for rendering scenes in the digital world. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Experimental results of attacking DUSt3R in the digital world with various background discs and view counts. Lower values of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of experimental results for discs with varying backgrounds across different camera pose estimation models in the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (a) Naturally symmetrical textures found in nature (denoted as N-1,2,3) and unoptimized radially symmetrical textures derived [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 41 canonical work pages

  1. [1]

    Seitz, and Richard Szeliski

    Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz, and Richard Szeliski. Building rome in a day. In Proceed- ings of IEEE International Conference on Computer Vision (ICCV), pages 72–79, 2009. 1

    work page 2009

  2. [2]

    Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation

    Jason Ansel, Edward Yang, Horace He, and et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Sup- port for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’24), 2024. 3

    work page 2024

  3. [3]

    Nature’s Patterns: A Tapestry in Three Parts

    Philip Ball. Nature’s Patterns: A Tapestry in Three Parts . Oxford University Press, 2009. 2

    work page 2009

  4. [4]

    G. Bradski. The opencv library. Dr. Dobb’s Journal of Soft- ware Tools, 2000. 2, 4

    work page 2000

  5. [5]

    Adversarial patch

    Tom B Brown, Dandelion Man ´e, Aurko Roy, Mart´ın Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. 1

    arXiv 2017

  6. [6]

    Towards evaluating the robustness of neural networks

    Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In Proceedings of IEEE Sym- posium on Security and Privacy (SP) , pages 39–57, 2017. 1

    work page 2017

  7. [7]

    Robust feature-level adversaries are in- terpretability tools

    Stephen Casper, Max Nadeau, Dylan Hadfield-Menell, and Gabriel Kreiman. Robust feature-level adversaries are in- terpretability tools. In Proceedings of Annual Conference on Neural Information Processing Systems (NeurIPS), pages 33093–33106, 2022. 1

    work page 2022

  8. [8]

    Shape matters: Deformable patch attack

    Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: Deformable patch attack. In Proceedings of European Conference on Computer Vision (ECCV), pages 529–548, 2022. 1

    work page 2022

Show all 44 references
  1. [9]

    Discrete-continuous optimization for large- scale structure from motion

    David Crandall, Andrew Owens, Noah Snavely, and Dan Huttenlocher. Discrete-continuous optimization for large- scale structure from motion. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001–3008, 2011. 1

    2011

  2. [10]

    Neural parametric gaussians for monocular non-rigid object reconstruction

    Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10715–10725, 2024. 1

    2024

  3. [11]

    X. Ding, J. Chen, H. Yu, Y . Shang, Y . Qin, and H. Ma. Transferable adversarial attacks for object detection using object-aware significant feature distortion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1546–1554, 2024. 1

    2024

  4. [12]

    Invisible pedestrians: Synthesiz- ing adversarial clothing textures to evade industrial camera- based 3d detection

    Xinlong Ding, Hongwei Yu, Jiansheng Chen, Jinlong Wang, Jintai Du, and Huimin Ma. Invisible pedestrians: Synthesiz- ing adversarial clothing textures to evade industrial camera- based 3d detection. In Proceedings of International Confer- ence on Multimedia and Expo (ICME), pag...

    2024

  5. [13]

    Boosting adversarial at- tacks with momentum

    Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial at- tacks with momentum. In Proceedings of IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pages 9185–9193, 2018. 1

    2018

  6. [14]

    Learning to produce semi-dense correspondences for visual localization

    Khang Truong Giang, Soohwan Song, and Sungho Jo. Learning to produce semi-dense correspondences for visual localization. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 19468–19478, 2024. 1

    2024

  7. [15]

    Explaining and harnessing adversarial examples

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1

    2014 arXiv

  8. [16]

    Poly haven: The public 3d asset library

    Poly Haven. Poly haven: The public 3d asset library. https://polyhaven.com/, 2024. 6

    2024

  9. [17]

    Z. Hu, S. Huang, X. Zhu, F. Sun, B. Zhang, and X. Hu. Ad- versarial texture for fooling person detectors in the physical world. In Proceedings of IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 13307– 13316, 2022. 1

    2022

  10. [18]

    T-sea: Transfer-based self-ensemble attack on object detection

    Hao Huang, Ziyan Chen, Huanran Chen, Yongtao Wang, and Kevin Zhang. T-sea: Transfer-based self-ensemble attack on object detection. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20514–20523, 2023. 1

    2023

  11. [19]

    Few-view object reconstruction with unknown cate- gories and camera poses

    Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown cate- gories and camera poses. In International Conference on 3D Vision (3DV), pages 31–41, 2024. 1

    2024

  12. [20]

    Efficient hyperparameter optimization with adaptive fidelity identification

    Jiantong Jiang, Zeyi Wen, Atif Mansoor, and Ajmal Mian. Efficient hyperparameter optimization with adaptive fidelity identification. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 26181–26190, 2024. 1

    2024

  13. [21]

    Lavan: Localized and visible adversarial noise

    Danny Karmon, Daniel Zoran, and Yoav Goldberg. Lavan: Localized and visible adversarial noise. InProceedings of In- ternational Conference on Machine Learning (ICML), pages 2507–2515, 2018. 1

    2018

  14. [22]

    Ad- versarial examples in the physical world

    Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Ad- versarial examples in the physical world. In Proceedings of International Conference on Learning Representations (ICLR), 2017. 1

    2017

  15. [23]

    Generalizable novel-view synthesis using a stereo camera

    Haechan Lee, Wonjoon Jin, Seung-Hwan Baek, and Sunghyun Cho. Generalizable novel-view synthesis using a stereo camera. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4939–4948, 2024. 1

    2024

  16. [24]

    Ground- ing image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Ground- ing image matching in 3d with mast3r. In Proceedings of European Conference on Computer Vision (ECCV), 2024. 1, 3, 5, 7

    2024

  17. [25]

    A²rnet: Adversarial attack resilient network for robust in- frared and visible image fusion

    Jiawei Li, Hongwei Yu, Jiansheng Chen, Xinlong Ding, Jin- long Wang, Jinyuan Liu, Bochao Zou, and Huimin Ma. A²rnet: Adversarial attack resilient network for robust in- frared and visible image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , p...

    2025

  18. [26]

    Relpose++: Recovering 6d poses from sparse-view ob- servations

    Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tul- siani. Relpose++: Recovering 6d poses from sparse-view ob- servations. In International Conference on 3D Vision (3DV), pages 106–115, 2024. 1, 5, 8

    2024

  19. [27]

    3d geometry-aware deformable gaussian splatting for dynamic view synthesis

    Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8900–8910, 2024. 1

    2024

  20. [28]

    Towards deep learning models resistant to adversarial attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of In- ternational Conference on Learning Representations (ICLR),

  21. [29]

    Accelerating 3d deep learning with pytorch3d

    Nikhila Ravi, Jeremy Reizenstein, David Novotny, Tay- lor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 2

    2007 arXiv

  22. [30]

    Chris Rockwell, Justin Johnson, and David F. Fouhey. The 8-point algorithm as an inductive bias for relative pose pre- diction by vits. In International Conference on 3D Vision (3DV), pages 1–11, 2022. 1

    2022

  23. [31]

    how do i organize my holiday snaps?

    Frederik Schaffalitzky and Andrew Zisserman. Multi-view matching for unordered image sets, or “how do i organize my holiday snaps?”. In Proceedings of European Conference on Computer Vision (ECCV), pages 414–431, 2002. 1

    2002

  24. [32]

    Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B

    Samarth Sinha, Jason Y . Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B. Lindell. Sparsepose: Sparse- view camera pose regression and refinement. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21349–21359, 2023. 1

    2023

  25. [33]

    In- triguing properties of neural networks

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. In- triguing properties of neural networks. In Proceedings of In- ternational Conference on Learning Representations (ICLR),

  26. [34]

    Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment

    Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9773–9783,

  27. [35]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697–20709, 2024. 1, 2, 3, 5, 7

    2024

  28. [36]

    Simul- taneously optimizing perturbations and positions for black- box adversarial patch attacks

    Xingxing Wei, Ying Guo, Jie Yu, and Bo Zhang. Simul- taneously optimizing perturbations and positions for black- box adversarial patch attacks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9041–9054, 2023. 1

    2023

  29. [37]

    X. Wei, J. Yu, and Y . Huang. Physically adversarial infrared patches with learnable shapes and locations. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12334–12342, 2023. 1

    2023

  30. [38]

    Robust global translations with 1dsfm

    Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Proceedings of European Conference on Computer Vision (ECCV), pages 61–75, 2014. 1

    2014

  31. [39]

    Physical adversarial attack on vehicle detector in the carla simulator

    Tong Wu, Xuefei Ning, Wenshuo Li, Ranran Huang, Huazhong Yang, and Yu Wang. Physical adversarial attack on vehicle detector in the carla simulator. arXiv preprint arXiv:2007.16118, 2020. 1

    2007 arXiv

  32. [40]

    Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation

    Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of IEEE/CVF Conferen...

    2023

  33. [41]

    Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, S ¨oren Pirk, and Arie E. Kaufman. Carve3d: Improving multi-view reconstruction consistency for diffu- sion models with rl finetuning. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2024

  34. [42]

    H. Yu, J. Chen, X. Ding, Y . Zhang, T. Tang, and H. Ma. Step vulnerability guided mean fluctuation adversarial at- tack against conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6791–6799, 2024. 1

    2024

  35. [43]

    Zhang, Deva Ramanan, and Shubham Tulsiani

    Jason Y . Zhang, Deva Ramanan, and Shubham Tulsiani. Rel- pose: Predicting probabilistic relative rotation for single ob- jects in the wild. In Proceedings of European Conference on Computer Vision (ECCV), page 592–611, 2022. 1

    2022

  36. [44]

    Cameras as rays: Pose estimation via ray diffusion

    Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. InProceedings of In- ternational Conference on Learning Representations (ICLR),

Pith tools

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