Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot Pose Estimation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that attaching pixel-wise uncertainty estimates to images generated by a multi-view diffusion prior lets a robot estimate a novel object's 6D pose and build a complete 3D model from a single RGB-D frame, refining both as n

desk verdict A genuinely useful integration of diffusion priors, uncertainty, and 3DGS for model-free pose estimation, but the uncertainty estimates that carry the method are never calibrated against actual reconstruction error. read the letter →

arxiv 2508.15972 v1 pith:YCEE6LRQ submitted 2025-08-21 cs.RO cs.CV

Zhaodong Jiang , Ashish Sinha , Tongtong Cao , Yuan Ren , Bingbing Liu , Binbin Xu This is my paper
classification cs.ROcs.CV
keywords 6Dposeestimationmodel-freezero-shotdiffusionpriorsepistemicuncertainty3DGaussianSplattinggraphoptimizationRGB-Dobjectreconstruction
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

UnPose tackles the last missing input in open-world 6D pose estimation: the object's own 3D model. Instead of requiring a CAD model, category training, or a set of camera-posed reference views, it starts from one RGB-D frame plus an object mask, and lets a pre-trained multi-view diffusion model synthesize the unseen sides of the object. The paper's central idea is that those synthesized views can only be trusted where the generator is confident, so it estimates a per-pixel epistemic uncertainty for every generated image and uses those uncertainties to weight both 3D Gaussian Splatting (3DGS) fusion and pose-graph optimization. As additional real observations arrive, high-uncertainty regions are corrected and low-uncertainty prior geometry is kept, yielding a complete, globally consistent model and an accurate 6D pose. The paper reports consistent improvements over model-based, zero-shot, and reconstruction-based baselines on standard benchmarks and shows the pipeline driving a real robot grasp of a mug after the handle becomes visible.

What carries the argument

The load-bearing machinery is an uncertainty-propagating denoising loop. A last-layer Laplace approximation gives a Gaussian predictive distribution for the diffusion model's noise prediction; a variance update rule carries that covariance through each deterministic denoising step, with the correlation between the noisy state and the prediction estimated by Monte Carlo sampling. The output is a per-pixel epistemic uncertainty map for every synthesized view. Those maps are then used twice: as pixel weights in the 3D Gaussian Splatting (3DGS—a scene represented by colored, semi-transparent Gaussian blobs) mapping loss, and as per-match weights in the geometric pose-graph residual, so that gene

What would settle it

Take a mug or another object with a heavily occluded hidden side, initialize from one RGB-D frame, and record the predicted per-pixel uncertainty on the hidden side; then reveal that side and compare it with the actual reconstruction error there. If the hidden side is not among the most uncertain pixels, or if the uncertainty does not decrease as matched observations accumulate, the uncertainty signal is not tracking the generator's error and cannot be the mechanism behind the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that epistemic uncertainty from a diffusion prior is what makes zero-shot model-free pose estimation reliable. UnPose applies a last-layer Laplace approximation to the diffusion model's noise prediction and propagates its covariance through the deterministic denoising update, estimating the intractable state-prediction covariance with Monte Carlo samples. These pixel-wise uncertainties weight the 3D Gaussian Splatting (3DGS) mapping loss and the feature matches in a pose graph that jointly optimizes real and diffusion-generated virtual keyframes. The paper argues this yields a complete, metrically consistent 3DGS model and accurate 6D pose from one RGB-D frame, improving

Load-bearing premise

The load-bearing premise—unverified quantitatively in the paper—is that the per-pixel uncertainties produced by last-layer Laplace approximation and Monte Carlo covariance propagation faithfully track the diffusion model's real error on unobserved geometry; if they are miscalibrated, the uncertainty-weighted losses have no principled basis and the improvements become a heuristic weighting of generated views.

Editorial extensions

If this is right

  • A robot with an RGB-D camera, object masks, and an off-the-shelf pose tracker can begin manipulating an unseen object after a single frame, with no CAD model or category training.
  • Pose accuracy and reconstruction quality improve incrementally as more observations arrive, so the estimate behaves as an online process rather than a one-shot reconstruction.
  • Diffusion-generated virtual views are most valuable when real observations are sparse; the paper's 8-image ablation attributes a 15 percent reconstruction gain to including them in the pose graph.
  • Removing uncertainty weighting or bundle adjustment degrades PSNR and Chamfer distance in the paper's ablations, indicating those mechanisms, not the diffusion prior alone, drive the reported gains.
  • The full initialization costs about 13.5 seconds, most of it Monte Carlo uncertainty estimation, so the method is slower than single-view baselines but much faster than optimization-based reconstruction methods.

Reading between the lines

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

  • The per-pixel uncertainties could double as an active next-best-view cue: a robot could steer its camera toward the most uncertain object regions, likely accelerating the accuracy gains the paper observes with added views.
  • If uncertainties were calibrated against actual reconstruction error, the same uncertainty-gated fusion could be applied to any image-to-3D generator, making this a general integration layer rather than a property of one specific diffusion model.
  • The method currently assumes object masks and tracked camera poses; combining it with open-vocabulary segmentation and a monocular SLAM would test how much of the zero-shot promise survives without those helping inputs.
  • A direct test on textureless objects, where the paper acknowledges the correspondence module struggles, would clarify whether the pose-graph benefits persist when geometric features carry little information.
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

4 major / 5 minor

Summary. UnPose proposes a zero-shot, model-free 6D pose estimation and 3D reconstruction pipeline from a single RGB-D frame. It uses Wonder3D to synthesize multi-view pseudo-images and applies a last-layer Laplace approximation (LLLA) to estimate pixel-wise epistemic uncertainty for the diffusion predictions. The generated views are aligned to the real frame via VGGT, PCA scale recovery, and ICP, and then fused into a 3D Gaussian Splatting field whose mapping loss is weighted by the estimated uncertainties. As more observations arrive, poses are refined in a pose graph that includes both real and diffusion-rendered frames, with diffusion-frame poses fixed. Pose estimation is obtained by running FoundationPose's refinement network on the reconstructed 3DGS field. Experiments on YCB-Video, LM-O, T-LESS, and TYO-L report improvements over GigaPose, SAM6D, and FoundationPose, and a real-robot manipulation demo is included.

Significance. The stated problem is important: estimating 6D pose and reconstructing unseen objects without CAD models is a key requirement for open-world robotics. The proposed pipeline is original in combining diffusion-based multi-view generation, LLLA uncertainty, 3DGS mapping, and pose-graph optimization. If the uncertainty estimates are well-calibrated, the uncertainty-guided weighting could be a genuine contribution. The paper also includes a real-world robotic demonstration, which strengthens the practical relevance. However, the central uncertainty mechanism is not validated against ground truth, and the ablation does not isolate the value of the specific LLLA-based uncertainties. The significance is therefore conditional on the uncertainty estimates being meaningfully correlated with reconstruction error, which the paper currently does not establish.

major comments (4)
  1. [Sec. 3.1, Eqs. (3)-(4)] Eq. (3) is presented as the variance update for the DDIM step, but no derivation is given. It appears to assume that the denoiser's dependence on x_t is locally linear and that higher-order correlations can be discarded; this is not justified. The cross-covariance term in Eq. (4) is estimated by Monte Carlo samples 'drawn around the current state x_t,' but the proposal distribution is not specified, and no convergence or accuracy check is provided. Since these uncertainty maps are the load-bearing element of the method and enter Eqs. (5)-(6), the paper needs to validate them, e.g., by comparing predicted uncertainty to actual per-pixel error or Chamfer distance in unobserved regions. Without such validation, the uncertainty-weighted losses reduce to heuristic weighting.
  2. [Table 1, Sec. 4.2] The ablation 'w/o Uncertainty Estimation' removes uncertainty entirely. It does not test whether the LLLA-based uncertainties carry information beyond any reasonable weighting scheme. A more informative ablation would replace C(p) with constant weights, inverse depth variance, or random weights, and show that the estimated uncertainties yield measurably better reconstruction. Additionally, Table 1 reports no error bars or standard deviations; given the stochastic nature of diffusion sampling and Monte Carlo estimation, single-run numbers are insufficient to support the claimed gains.
  3. [Sec. 3.4, last paragraph] The paper states that 'we fix the poses of all diffusion-rendered frames' in the pose graph optimization. This means systematic errors in the VGGT/PCA/ICP alignment of the diffusion frames can never be corrected, and the final reconstruction is biased toward the diffusion prior. The paper should quantify the alignment error of the diffusion frames against ground truth and test the sensitivity to this choice, e.g., by also optimizing virtual frame poses with a strong regularization term.
  4. [Sec. 4.1 and Tables 4-6] The main-text claim that UnPose 'surpasses GigaPose by ������ and FoundationPose by �����' contains literal blank placeholders, so the headline quantitative comparison is missing in the main body. The appendix tables do contain numbers, but these are reported without error bars or multiple-run statistics. Since the diffusion prior and Monte Carlo sampling introduce randomness, the reported ADD/ADD-S and Chamfer distance values should be accompanied by variance or confidence intervals.
minor comments (5)
  1. [Sec. 3.1, first paragraph] The number of Monte Carlo samples M is given as '� ��' (a placeholder) in the text; the actual value should be stated explicitly. Also, the sentence 'we use ... samples per step across 50 denoising steps' should specify whether this is M=??.
  2. [Eq. (5)] The mapping loss is not fully specified: the norms for the depth and color residuals are not defined (L1 vs L2), and the visibility score S(p) is not described in detail. Please clarify.
  3. [Fig. 3 and Fig. 9] The captions say 'larger variance at unseen perspective' — consider rewording to 'unseen viewpoints' for clarity.
  4. [Table 3] For Wonder3D, the 8- and 16-image columns are all zeros because the method is single-view only; this should be noted explicitly rather than leaving empty/zero entries that may be misinterpreted.
  5. [General presentation] The LaTeX rendering has numerous '�' placeholders throughout the text (e.g., Sec. 4.1 and Sec. 3.1). These must be fixed before a final submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central pose and reconstruction claims are evaluated against external ground truth; the uncertainty weighting is a heuristic robustness choice, not a definitional reduction.

full rationale

The paper's central claims—6D pose accuracy and 3D reconstruction quality—are validated against external ground-truth poses and Chamfer distances (e.g., Tables 3–6, Figs. 5–7), not against the diffusion model's own outputs or its uncertainty estimates. The uncertainty estimates from Eqs. (3)–(4) are used as weights in the mapping loss (Eq. 5) and pose-graph residuals (Eq. 6), but the predicted quantities are not defined as functions of those uncertainties. Using a model's self-reported confidence to weight its own generated views is a heuristic that may be miscalibrated, and the paper does not calibrate the uncertainty against external reconstruction error; however, that is a validity/robustness concern, not a circular argument under the specified criteria. The paper's only potentially self-referential citation is [22] (Hippo), which shares authors with this work, but it appears in Related Work as background and is not used to justify a uniqueness theorem or to supply a load-bearing premise. The pose-graph design choice to fix diffusion-rendered frames is a stated implementation detail, not the quantity being predicted. The limitations section explicitly acknowledges dependence on Mast3R correspondences for textureless objects and the computational cost of Monte Carlo sampling, which further confirms that the authors are not concealing a circular reduction. Overall, no derivation step reduces by construction to its own inputs, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new physical entities. Its load-bearing assumptions are (a) the validity of the uncertainty propagation through DDIM (an ad hoc approximation), (b) the reliability of the pre-trained components (Wonder3D, VGGT, FoundationPose, Mast3R, ORB-SLAM3), and (c) the correct alignment of generated views via PCA+ICP. Several hyperparameters (n=6, M=16, 50 steps) are chosen by hand without sensitivity analysis, and the keyframe threshold is unspecified.

free parameters (5)
  • number of diffusion views n = 6
    Set in Sec. 3.1 with no sensitivity analysis; affects the amount of prior information injected.
  • Monte Carlo samples M per denoising step = 16
    Chosen in Sec. 3.1 to balance quality and cost; no ablation on the number of samples.
  • number of DDIM denoising steps = 50
    Used in Sec. 3.1 for uncertainty estimation; standard but not justified for this task.
  • keyframe matching inlier threshold
    Required in Sec. 3.4 for deciding when to add a keyframe; not specified numerically.
  • per-match weighting in Eq. (6)
    The extended per-match weighting with diffusion uncertainties is not given numeric form or parameters.
assumptions (7)
  • domain assumption Camera poses are known via an off-the-shelf SLAM method (ORB-SLAM3).
    Stated in Sec. 3: 'Camera poses, T_{wc} in SE(3), are known using an off-the-shelf SLAM method [45].' The entire multi-view fusion relies on this.
  • domain assumption Object masks M are provided for all objects in the scene.
    Sec. 3.1: 'We assume that object masks M are provided to us beforehand for n objects in the scene using a SOTA instance segmentation model.'
  • domain assumption Wonder3D generates multiview images that are useful priors for unseen object geometry.
    The whole method depends on the pre-trained Wonder3D producing realistic novel views; no failure modes are analyzed beyond the uncertainty weighting.
  • domain assumption Last-layer Laplace approximation yields a valid predictive covariance for the diffusion noise predictor.
    Sec. 3.1 assumes Eq. (1), a Gaussian approximation of the noise prediction distribution, following BayesDiff/LLLA.
  • ad hoc to paper The Monte Carlo covariance estimate in Eq. (4) approximates the true correlation between noisy state and model prediction.
    The covariance term Cov(x_t, epsilon_theta) is declared 'analytically intractable' and replaced with a Monte Carlo formula that is not derived or validated.
  • domain assumption PCA eigenvalue matching with ICP recovers the correct scale and alignment between diffusion views and real-world coordinates.
    Sec. 3.1: 'we first match the eigenvalues of their covariance matrices via PCA... After recovering the scale s, we further refine the rigid transformation between them using ICP.'
  • domain assumption FoundationPose's pose refinement network transfers from mesh-rendered inputs to 3DGS-rendered inputs.
    Sec. 3.3: 'we modify the first input to be rendered directly from the 3DGS field, rather than a pre-reconstructed 3D model.' No adaptation or retraining is done.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot Pose Estimation." pith.science (2026). https://pith.science/paper/YCEE6LRQ

@misc{pith2026250815972,
  author       = {Pith},
  title        = {Pith review of: UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot Pose Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCEE6LRQ}},
  note         = {Machine review of arXiv:2508.15972}
}
read the original abstract

Estimating the 6D pose of novel objects is a fundamental yet challenging problem in robotics, often relying on access to object CAD models. However, acquiring such models can be costly and impractical. Recent approaches aim to bypass this requirement by leveraging strong priors from foundation models to reconstruct objects from single or multi-view images, but typically require additional training or produce hallucinated geometry. To this end, we propose UnPose, a novel framework for zero-shot, model-free 6D object pose estimation and reconstruction that exploits 3D priors and uncertainty estimates from a pre-trained diffusion model. Specifically, starting from a single-view RGB-D frame, UnPose uses a multi-view diffusion model to estimate an initial 3D model using 3D Gaussian Splatting (3DGS) representation, along with pixel-wise epistemic uncertainty estimates. As additional observations become available, we incrementally refine the 3DGS model by fusing new views guided by the diffusion model's uncertainty, thereby continuously improving the pose estimation accuracy and 3D reconstruction quality. To ensure global consistency, the diffusion prior-generated views and subsequent observations are further integrated in a pose graph and jointly optimized into a coherent 3DGS field. Extensive experiments demonstrate that UnPose significantly outperforms existing approaches in both 6D pose estimation accuracy and 3D reconstruction quality. We further showcase its practical applicability in real-world robotic manipulation tasks.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Uncertainty-Calibrated Diffusion for Reliable 3D Molecular Graph Generation

    cs.LG 2026-06 unverdicted novelty 4.0 of 10

    UCD adjusts diffusion-based 3D molecular graph generation to handle epistemic uncertainty, improving sample quality and reaching new benchmark performance.

Reference graph

Works this paper leans on

63 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1, 2023. 1, 2, 3, 4, 5, 6

    work page 2023

  2. [2]

    Zhou, Y .-S

    J. Zhou, Y .-S. Liu, and Z. Han. Zero-shot scene reconstruction from single images with deep prior assembly. In Advances in Neural Information Processing Systems (NeurIPS) , 2024. 2

    work page 2024

  3. [3]

    Sundermeyer, A

    M. Sundermeyer, A. Mousavian, R. Triebel, and D. Fox. Contact-graspnet: Efficient 6-dof grasp gen- eration in cluttered scenes. IEEE International Conference on Robotics and Automation (ICRA) , 2021. 2

    work page 2021

  4. [4]

    K. Park, T. Patten, and M. Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision , pages 7668– 7677, 2019. 2

    work page 2019

  5. [5]

    Labb´e, J

    Y . Labb´e, J. Carpentier, M. Aubry, and J. Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16, pages 574–591. Springer, 2020. 2

    work page 2020

  6. [6]

    Y . He, W. Sun, H. Huang, J. Liu, H. Fan, and J. Sun. Pvn3d: A deep point-wise 3d keypoints voting network for 6dof pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11632–11641, 2020. 2

    work page 2020

  7. [7]

    Y . He, H. Huang, H. Fan, Q. Chen, and J. Sun. Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3003–3013, 2021. 2

    work page 2021

  8. [8]

    H. Wang, S. Sridhar, J. Huang, J. Valentin, S. Song, and L. J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2642–2651, 2019. 2

    work page 2019

Show all 63 references
  1. [9]

    Zhang, W

    J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, and H. Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199–216. Springer, 2024. 2

    2024

  2. [10]

    D. Chen, J. Li, Z. Wang, and K. Xu. Learning canonical shape space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11973–11982, 2020. 2

    2020

  3. [11]

    X. Lin, W. Yang, Y . Gao, and T. Zhang. Instance-adaptive and geometric-aware keypoint learning for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21040–21049, 2024. 2

    2024

  4. [12]

    Y . Chen, Y . Di, G. Zhai, F. Manhardt, C. Zhang, R. Zhang, F. Tombari, N. Navab, and B. Busam. Sec- ondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ,...

    2024

  5. [13]

    Caraffa, D

    A. Caraffa, D. Boscaini, A. Hamza, and F. Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. In European Conference on Computer Vision , pages 414–431. Springer, 2024. 2

    2024

  6. [14]

    J. Lin, L. Liu, D. Lu, and K. Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 27906–27916, 2024. 2, 3, 7, 13, 14, 15, 16

    2024

  7. [15]

    J. Chen, Z. Zhou, M. Sun, R. Zhao, L. Wu, T. Bao, and Z. He. Zeropose: Cad-prompted zero-shot object 6d pose estimation in cluttered scenes. IEEE Transactions on Circuits and Systems for Video Technology,

  8. [16]

    J. Sun, Z. Wang, S. Zhang, X. He, H. Zhao, G. Zhang, and X. Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825–6834, 2022. 2

    2022

  9. [17]

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou. Loftr: Detector-free local feature matching with trans- formers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8922–8931, 2021. 2, 3

    2021

  10. [18]

    V . N. Nguyen, T. Groueix, G. Ponimatkin, Y . Hu, R. Marlet, M. Salzmann, and V . Lepetit. Nope: Novel object pose estimation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17923–17932, 2024. 2, 3

    2024

  11. [19]

    Y . Liu, Y . Wen, S. Peng, C. Lin, X. Long, T. Komura, and W. Wang. Gen6d: Generalizable model- free 6-dof object pose estimation from rgb images. In European Conference on Computer Vision , pages 298–315. Springer, 2022. 2

    2022

  12. [20]

    T. Lee, B. Wen, M. Kang, G. Kang, I. S. Kweon, and K.-J. Yoon. Any6d: Model-free 6d pose estimation of novel objects. CVPR, 2025. 2, 3, 14

    2025

  13. [21]

    V . N. Nguyen, T. Groueix, M. Salzmann, and V . Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2, 3, 7, 13, 14, 15, 16

    2024

  14. [22]

    Y . Liu, Z. Jiang, B. Xu, G. Wu, Y . Ren, T. Cao, B. Liu, R. H. Yang, A. Rasouli, and J. Shan. Hippo: Har- nessing image-to-3d priors for model-free zero-shot 6d pose estimation. IEEE Robotics and Automation Letters, 10(8):8284–8291, 2025. 2, 3, 14

    2025

  15. [23]

    Kendall and Y

    A. Kendall and Y . Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems , 30, 2017. 2, 3

    2017

  16. [24]

    Thrun, W

    S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics. Cambridge: MIT Press, 2005. 2, 3

    2005

  17. [25]

    Z. Liao, B. Xu, and S. L. Waslander. Toward general object-level mapping from sparse views with 3d diffusion priors. In 8th Annual Conference on Robot Learning , 2024. 2, 3, 7, 14

    2024

  18. [26]

    D. Cai, J. Heikkil ¨a, and E. Rahtu. Gs-pose: Cascaded framework for generalizable segmentation-based 6d object pose estimation. arXiv e-prints, pages arXiv–2403, 2024. 2

    2024

  19. [27]

    Ponimatkin, M

    G. Ponimatkin, M. C ´ıfka, T. Souˇcek, M. Fourmy, Y . Labb´e, V . Petrik, and J. Sivic. 6d object pose tracking in internet videos for robotic manipulation. arXiv preprint arXiv:2503.10307, 2025. 2, 3

    2025 arXiv

  20. [28]

    X. Liu, G. Wang, R. Zhang, C. Zhang, F. Tombari, and X. Ji. Unopose: Unseen object pose estimation with an unposed rgb-d reference image. arXiv preprint arXiv:2411.16106, 2024. 3

    2024 arXiv

  21. [29]

    B. Wen, J. Tremblay, V . Blukis, S. Tyree, T. M ¨uller, A. Evans, D. Fox, J. Kautz, and S. Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 606–61...

    2023

  22. [30]

    B. Wen, W. Yang, J. Kautz, and S. Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868–17879, 2024. 3, 6, 7, 13, 14, 15, 16

    2024

  23. [31]

    Corsetti, D

    J. Corsetti, D. Boscaini, C. Oh, A. Cavallaro, and F. Poiesi. Open-vocabulary object 6d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18071– 18080, 2024. 3 10

    2024

  24. [32]

    S. K. Aithal, P. Maini, Z. Lipton, and J. Z. Kolter. Understanding hallucinations in diffusion models through mode interpolation. Advances in Neural Information Processing Systems , 37:134614–134644,

  25. [33]

    McCormac, R

    J. McCormac, R. Clark, M. Bloesch, A. Davison, and S. Leutenegger. Fusion++: V olumetric object-level slam. In 2018 international conference on 3D vision (3DV) , pages 32–41, 2018. 3

    2018

  26. [34]

    B. Xu, W. Li, D. Tzoumanikas, M. Bloesch, A. Davison, and S. Leutenegger. Mid-fusion: Octree-based object-level multi-instance dynamic slam. In 2019 International Conference on Robotics and Automation (ICRA), 2019. 3

    2019

  27. [35]

    Nicholson, M

    L. Nicholson, M. Milford, and N. S ¨underhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. IEEE Robotics and Automation Letters , 4(1):1–8, 2018. 3

    2018

  28. [36]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021. 3

    2021

  29. [37]

    Dellaert, M

    F. Dellaert, M. Kaess, et al. Factor graphs for robot perception. F oundations and Trends® in Robotics, 6 (1-2):1–139, 2017. 3

    2017

  30. [38]

    Z. Liao, J. Yang, J. Qian, A. P. Schoellig, and S. L. Waslander. Uncertainty-aware 3d object-level mapping with deep shape priors. 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024. 3

    2024

  31. [39]

    Z. Tan, X. Chen, J. Zhang, L. Feng, and D. Hu. Uncertainty-aware normal-guided gaussian splatting for surface reconstruction from sparse image sequences. arXiv preprint arXiv:2503.11172, 2025. 3

    2025 arXiv

  32. [40]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems , 30, 2017. 3

    2017

  33. [41]

    L. V . Jospin, H. Laga, F. Boussaid, W. Buntine, and M. Bennamoun. Hands-on bayesian neural net- works—a tutorial for deep learning users. IEEE Computational Intelligence Magazine , 17(2):29–48,

  34. [42]

    Blundell, J

    C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight uncertainty in neural network. In International conference on machine learning , pages 1613–1622. PMLR, 2015. 3

    2015

  35. [43]

    Ritter, A

    H. Ritter, A. Botev, and D. Barber. A scalable laplace approximation for neural networks. In 6th in- ternational conference on learning representations, ICLR 2018-conference track proceedings , volume 6. International Conference on Representation Learning, 2018. 3

    2018

  36. [44]

    S. Kou, L. Gan, D. Wang, C. Li, and Z. Deng. Bayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference. ICLR, 2024. 4

    2024

  37. [45]

    Campos, R

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os. Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. IEEE Transactions on Robotics , 37 (6):1874–1890, 2021. 4

    2021

  38. [46]

    Long, Y .-C

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970–9980, 2024. ...

    2024

  39. [47]

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny. Vggt: Visual geometry grounded transformer. arXiv preprint arXiv:2503.11651, 2025. 4, 5

    2025 arXiv

  40. [48]

    Kristiadi, M

    A. Kristiadi, M. Hein, and P. Hennig. Being bayesian, even just a bit, fixes overconfidence in relu net- works. In International conference on machine learning , pages 5436–5446. PMLR, 2020. 4

    2020

  41. [49]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. ICLR, 2021. 4

    2021

  42. [50]

    W. K. Hastings. Monte Carlo sampling methods using Markov chains and their applications . Oxford University Press, 1970. 5

    1970

  43. [51]

    Zhang, Y

    J. Zhang, Y . Yao, and B. Deng. Fast and robust iterative closest point. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3450–3466, 2021. 5

    2021

  44. [52]

    Keetha, J

    N. Keetha, J. Karhade, K. M. Jatavallabhula, G. Yang, S. Scherer, D. Ramanan, and J. Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357–21366, 2024. 5 11

    2024

  45. [53]

    Xiang, T

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 6, 7, 8, 13, 14, 16

    2017 arXiv

  46. [54]

    Brachmann, A

    E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13 , pages 536–551....

    2014

  47. [55]

    Leroy, Y

    V . Leroy, Y . Cabon, and J. Revaud. Grounding image matching in 3d with mast3r. InEuropean Conference on Computer Vision, pages 71–91, 2024. 6, 9

    2024

  48. [56]

    Tolias, Y

    G. Tolias, Y . Avrithis, and H. J´egou. To aggregate or not to aggregate: Selective match kernels for image search. In IEEE International Conference on Computer Vision , 2013. 6

    2013

  49. [57]

    Murai, E

    R. Murai, E. Dexheimer, and A. J. Davison. MASt3R-SLAM: Real-time dense SLAM with 3D recon- struction priors. CVPR, 2025. 6

    2025

  50. [58]

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 7

    2021 arXiv

  51. [59]

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics (T-RO), 2023. 8, 14

    2023

  52. [60]

    Whelan, R

    T. Whelan, R. F. Salas-Moreno, B. Glocker, A. J. Davison, and S. Leutenegger. Elasticfusion: Real- time dense slam and light source estimation. The International Journal of Robotics Research , 35(14): 1697–1716, 2016. 9

    2016

  53. [61]

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 13

    2017

  54. [62]

    Hodan, P

    T. Hodan, P. Haluza, ˇS. Obdrˇz´alek, J. Matas, M. Lourakis, and X. Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In WACV, 2017. 13, 14

    2017

  55. [63]

    Hodan, F

    T. Hodan, F. Michel, E. Brachmann, W. Kehl, A. GlentBuch, D. Kraft, B. Drost, J. Vidal, S. Ihrke, X. Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In ECCV, pages 19–34, 2018. 13, 14 12 A Additional Qualitative Results Scene-level Reconstructions. To demonstrate...

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.