Pith. sign in

REVIEW 3 major objections 5 minor 61 references

Articulated Object Estimation in the Wild

T0 review · 3 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 ArtiPoint claims articulated object models—joint type, axis direction, and axis position for prismatic and revolute joints—can be estimated from egocentric RGB-D video of a human interacting with objects, even with a moving camera and parti

desk verdict ArtiPoint and Arti4D are a useful, well-engineered contribution to in-the-wild articulation estimation, but the 'raw RGB-D' claim needs to wait for an unmatched-interaction recall analysis and a stated odometry dependency. read the letter →

arxiv 2509.01708 v1 pith:N26WOS2A submitted 2025-09-01 cs.RO cs.CV

Abdelrhman Werby , Martin Büchner , Adrian Röfer , Chenguang Huang , Wolfram Burgard , Abhinav Valada This is my paper · ORCID
classification cs.ROcs.CV
keywords articulatedobjectestimationpointtrackingfactorgraphoptimizationRGB-Dvideoegocentricinteractionprismaticandrevolutejointsin-the-wilddataset3Dsceneunderstanding
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

ArtiPoint aims to infer the kinematic model of an articulated object—whether a drawer slides, a door hinges, and where the axis sits—from ordinary handheld RGB-D video of someone manipulating it. It does this by detecting the hand, tracking image points near it, lifting those trajectories into 3D, cancelling the camera motion, and fitting a single articulation twist to all point tracks with a factor graph. On the new Arti4D dataset, it reports prismatic axis errors around 14.5 degrees, revolute axis errors around 17.1 degrees, revolute axis position errors of 0.07 meters, and 68%/98% prismatic/revolute type accuracy, outperforming classical and learning-based baselines. If this holds, robots could learn to manipulate cabinets, drawers, and appliances by watching a human demonstration captured while walking around a room. The headline numbers, however, assume accurate camera poses are already known.

What carries the argument

The key mechanism is an articulation-regularized factor graph over point trajectories. A one-degree-of-freedom articulation is represented as a twist ξ = (ω, v), a six-parameter screw motion; moving through configuration θ applies the rigid transform exp(ξθ). Instead of estimating independent inter-frame transformations and then fitting a joint, ArtiPoint solves for one shared twist and all per-frame configuration parameters simultaneously, so every tracked point votes on the same axis. The input trajectories come from a deep any-point tracker applied to keypoints sampled around a detected hand, and the factor graph optimization is solved with a general-purpose factor graph solver.

What would settle it

Run ArtiPoint on Arti4D sequences after adding controlled noise to the provided ground-truth camera poses—for instance, translation noise from 1 to 5 cm and rotation noise from 1 to 5 degrees—and plot prismatic and revolute axis errors against the noise level. If errors grow sharply or joint-type accuracy collapses before the object motion itself becomes ambiguous, the claim that articulation estimation works under dynamic camera motion would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that articulated object parameters can be estimated directly from raw RGB-D videos of human-object interactions, without requiring a fixed camera, isolated objects, or full observability. ArtiPoint detects the manipulating hand to isolate interaction segments, samples query points in the hand's vicinity, tracks those points through the segment with a deep any-point tracker, lifts the tracks to 3D using depth, compensates for camera motion, and filters out static and unreliable tracks. It then estimates a single articulation model—joint type, axis direction, and axis position—by solving an articulation-regularized factor graph over all point trajectories. On the A

Load-bearing premise

The pipeline assumes the camera's position and orientation are already known accurately for every frame; the headline numbers come from external trackers, and if that pose information is noisy or absent the estimated axes inherit the error.

Editorial extensions

If this is right

  • Fixed-camera assumptions in prior articulation estimators can be dropped: a hand-held moving sensor with partial occlusion is sufficient.
  • Hand detection can automatically isolate interaction segments in unconstrained scenes, enabling extraction of demonstration data without manual labeling.
  • The same pipeline handles both prismatic and revolute joints category-agnostically, so drawers, doors, fridges, and sliders are covered by one method.
  • Arti4D provides ground-truth axes, interaction windows, difficulty labels, and camera poses, making it a benchmark for both articulation estimation and visual odometry in scenes with articulated motion.
  • If paired with reliable odometry, the approach supports learning manipulation skills from human demonstrations in everyday environments.

Reading between the lines

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

  • The dependence on accurate external camera poses could be dissolved by treating camera pose as unknowns in the same factor graph, jointly estimating camera motion and articulation axes; that would make the 'raw RGB-D' claim self-contained.
  • Swapping the hand-detection trigger for egocentric action recognition or gaze could extend the method to tool-mediated interaction, remote manipulation, or scenes where hands are not visible.
  • Because the twist representation covers screw motions, ArtiPoint may already be capable of estimating screw joints even though the experiments only report prismatic and revolute types; a dedicated screw-motion benchmark would test this.
  • A natural next step, not implemented in the paper, is to feed the estimated axes and part trajectories into a manipulation policy so a robot can replay the demonstrated articulation.
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 / 5 minor

Summary. The paper proposes ArtiPoint, a four-stage pipeline for estimating 1-DoF articulated object models from egocentric RGB-D video. The method extracts interaction intervals using a hand-detection trigger, samples and tracks keypoints near the hand with MobileSAM and CoTracker3, lifts and filters 3D tracks, and then estimates articulation axes and joint types through a factor-graph optimization that extends the estimator of Buchanan et al. with an articulation-regularized trajectory objective (Eq. 4). The authors introduce Arti4D, a 45-sequence, 414-interaction ego-centric dataset with ground-truth camera poses, interaction windows, and articulated-axis labels, and report that ArtiPoint outperforms ArtGS, Ditto, Sturm et al., and Buchanan et al. baselines on that dataset. The core claim is that ArtiPoint can infer articulated object models directly from raw RGB-D videos under dynamic camera motion and partial observability.

Significance. If the reported results hold, ArtiPoint would be a meaningful step toward in-the-wild articulated object estimation, where prior work has largely assumed static cameras, isolated objects, or full observability. The Arti4D dataset itself is a notable contribution and is released with code, model predictions, and ground-truth poses. The paper has several concrete strengths: the supplementary evaluation with DROID-SLAM poses (Table S.1) addresses, at least partially, the camera-odometry dependence; the ablations in Tables 3 and S.3 cover key components and hyperparameters; and the articulation-regularized estimator in Eq. (4) is a sensible extension that shows consistent improvement over independent transforms in Table 1. However, the empirical support for the headline 'raw RGB-D' claim is incomplete. The evaluation protocol matches predicted interaction segments to ground-truth segments and only scores matched predictions, but it never reports how many of the 414 interactions are missed by the hand-trigger segmentation front-end. Without a recall/failure analysis, Table 1 can materially overstate end-to-end performance. The absence of error bars or statistical tests also makes the sm

major comments (3)
  1. [Section S.1.1, Table 1, Sec. 4] The evaluation protocol computes all metrics only for predicted interaction segments with IoU > 0.5 to a ground-truth segment. The paper claims in Sec. 4 that the aim is to 'detect all interactions and estimate the articulation parameters of all manipulated objects,' and Sec. 5 says results are 'averaged across all object interactions,' but no recall, precision, or unmatched-interaction count is reported anywhere. The supplementary even admits in S.1.6 that ArtiPoint 'detects the majority of interactions' and in S.1.5 that RH078 contains non-separable interactions. If a substantial fraction of the 414 interactions are dropped because the hand is not detected or segments are merged/split, those failures never enter Table 1. This is load-bearing for the central claim that the method works 'directly from raw RGB-D videos.' The authors should report the number of predicted segments, the matc
  2. [Sec. 3.3, Table 1, Section S.1.2] The headline numbers in Table 1 are produced with ground-truth VIVE-tracker camera poses, while Sec. 3.3 states that the method 'transform[s] all the 3D point tracks ... assuming accurate camera odometry.' The supplementary shows that classical SLAM methods fail on the interaction segments and only DROID-SLAM gives usable poses, and even then the DROID-SLAM trajectories are registered to the ground-truth map with KISS-Matcher before evaluation. Table S.1 shows only a mild degradation, which is encouraging, but this is one SLAM method and the registration step may hide alignment errors. The phrase 'directly from raw RGB-D videos' in the abstract is therefore not fully supported by the main evaluation. The main protocol should either use estimated poses or the authors should prominently state that the principal results assume known camera poses and report the SLAM-based results as the raw-
  3. [Tables 1-3 and S.1-S.4] All quantitative claims of superiority rest on single mean values with no error bars, confidence intervals, or significance tests. Several margins are small: for example, ArtiPoint versus ArtiPoint with independent transforms in Table 1 gives prismatic θerr 14.54° vs 15.60°, revolute θerr 17.14° vs 18.61°, and revolute type accuracy 0.98 vs 0.96. The per-scene results in Table S.4 show large variability (e.g., prismatic θerr ranges from 7.86° in DR080 to 35.11° in RH078), so the overall means may be driven by specific scenes. The authors should report variances or per-segment distributions and, where possible, paired comparisons, especially for the ablation claims in Table 3.
minor comments (5)
  1. [Eqs. (2)-(4)] The temporal offset ε used in the observation pairs is never defined. Please state its value or explain how it is chosen; this affects the construction of the pose observations and the articulation-regularized objective.
  2. [Sec. 5.1, Table 3 caption] The text says 'revolute and rotational joints' in Sec. 5.3; this is likely a typo for 'revolute.' Also, Table 3 reports dL2 in cm while Table 1 uses m; please use consistent units.
  3. [Sec. 3.1/S.1.4.1] The values of τh and σstatic are not reported, although they are central to interaction extraction and static-track filtering. The paper gives defaults for wh, Tmin, Tmax, σreliable, λvel, and λjerk, but these two thresholds are omitted. Please add them and, if possible, show sensitivity in the supplementary.
  4. [Sec. 3.4] Minor wording: 'an extend the estimation' should be 'and extend the estimation.' Also, the relationship between the twist estimated by Eq. (4) and the final twist returned by the Buchanan et al. estimator is not entirely explicit; a sentence clarifying that Eq. (4) is a front-end to the final factor-graph estimator would improve reproducibility.
  5. [S.1.2] The DROID-SLAM poses are registered against the ground-truth point cloud with KISS-Matcher before evaluation. This should be described as a form of alignment that may partly compensate for global drift, so the results in Table S.1 should be interpreted with that caveat.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the articulation objective uses only point tracks and visibility, the Buchanan et al. back-end is independent published work, and the evaluation caveats are correctness issues rather than circularity.

full rationale

The derivation chain is self-contained. The articulation parameters are estimated by solving Eq. (4) from point tracks and visibility masks; ground-truth axes are not used in the objective and are only used for evaluation. The final twist is produced by the Buchanan et al. factor-graph estimator [9], a prior published method. Although two authors of this paper are also co-authors of [9], the citation is not circular: [9] is an externally falsifiable estimator, and the key comparisons in Table 1 run the same back-end under different front-ends, so the comparison isolates the contribution of the ArtiPoint front-end and articulation-regularized trajectory optimization. The main caveats are evaluation-selection and input-dependence issues, not circularity: Sec. S.1.1 matches predictions to ground-truth segments at IoU>0.5 and does not report recall of unmatched interactions, and Sec. S.1.2 states that the headline numbers use ground-truth VIVE odometry, with DROID-SLAM poses giving slightly worse results. The limitations in Sec. 7 (hand-detection trigger, two-body kinematics, depth dependence) and S.1.5 (non-separable interactions in RH078) are acknowledged and affect real-world robustness, but they do not make any equation or prediction equivalent to its input by construction. No circular step was found.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The pipeline introduces no new physical entities and no fitted constants of nature. It depends on several hand-chosen thresholds and smoothing weights, some of which are not fully specified in the text, and on domain assumptions about hands, depth quality, camera odometry, and 1-DoF kinematics. The external anchor of the claims is the Arti4D benchmark rather than a first-principles derivation.

free parameters (8)
  • tau_h (hand detection trigger threshold)
    Threshold on smoothed hand-mask signal for starting and ending interaction segments (Sec. 3.1); value not reported in the text, selected by hand.
  • Hand-detection window wh = 6
    Moving-average window for hand detections; default used in all reported results (Sec. 3.1, Sec. S.1.4.1).
  • Segment length bounds Tmin, Tmax = Tmin=30, Tmax=90
    Filters extracted interaction segments; chosen as defaults and ablated in Sec. S.1.4.1.
  • Static-track percentile threshold sigma_static
    Tracks with positional variance below a percentile are discarded (Sec. 3.3); the percentile value is not specified.
  • Reliability threshold sigma_reliable = 0.5
    Maximum allowed occlusion fraction for keeping a track (Sec. 3.3 and Implementation Details).
  • Smoothing weights lambda_vel, lambda_jerk = lambda_vel=0.5, lambda_jerk=5
    Weights in the trajectory smoothing objective Eq. (1), set in Implementation Details.
  • Keyframe stride for point tracking = 2 (implied)
    Defaults to stride 2 based on supplementary Fig. S.2; no explicit default is given in the main text.
  • Sampling density / number of keypoints F
    Number of GFTT query points per object mask is not specified; affects tracker density and later estimation.
assumptions (7)
  • standard math 1-DoF articulated motion is exactly representable as a twist xi in se(3) times configuration theta, with exp map to SE(3).
    Used in Sec. 3.4 to define both the pose extraction Eq. (3) and the articulation-regularized estimator Eq. (4).
  • domain assumption Hand presence is a reliable trigger for interaction segments.
    Sec. 3.1 segments video solely from hand-mask detections; the authors' limitation section notes this trigger requirement.
  • domain assumption Accurate camera odometry is available for each segment.
    Sec. 3.3 transforms tracks to a global frame using camera odometry; main experiments use ground-truth poses from Arti4D, not a raw-RGB-D-only estimation.
  • domain assumption Depth measurements and camera intrinsics correctly lift tracked 2D points to 3D.
    Sec. 3.3 relies on depth Dt and intrinsics K, filtering invalid depth values; reflective, metallic, or dark objects break this assumption per Sec. 7 and Fig. S.6.
  • domain assumption CoTracker3 visibility and tracking are reliable enough for downstream estimation.
    The front-end accepts CoTracker3 tracks and visibility V as observations (Sec. 3.2, Eq. (2)); no uncertainty from the tracker is propagated.
  • domain assumption Objects have two-body, 1-DoF articulation models.
    The se(3) twist model and factor graph assume single prismatic, revolute, or screw joints; the limitation section states multi-DoF articulation is infeasible.
  • domain assumption The factor graph back-end of Buchanan et al. [9] correctly recovers a twist from relative pose observations.
    Adopted as the final estimator in Sec. 3.4; correctness is inherited from the cited paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Articulated Object Estimation in the Wild." pith.science (2026). https://pith.science/paper/N26WOS2A

@misc{pith2026250901708,
  author       = {Pith},
  title        = {Pith review of: Articulated Object Estimation in the Wild},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N26WOS2A}},
  note         = {Machine review of arXiv:2509.01708}
}
read the original abstract

Understanding the 3D motion of articulated objects is essential in robotic scene understanding, mobile manipulation, and motion planning. Prior methods for articulation estimation have primarily focused on controlled settings, assuming either fixed camera viewpoints or direct observations of various object states, which tend to fail in more realistic unconstrained environments. In contrast, humans effortlessly infer articulation by watching others manipulate objects. Inspired by this, we introduce ArtiPoint, a novel estimation framework that can infer articulated object models under dynamic camera motion and partial observability. By combining deep point tracking with a factor graph optimization framework, ArtiPoint robustly estimates articulated part trajectories and articulation axes directly from raw RGB-D videos. To foster future research in this domain, we introduce Arti4D, the first ego-centric in-the-wild dataset that captures articulated object interactions at a scene level, accompanied by articulation labels and ground-truth camera poses. We benchmark ArtiPoint against a range of classical and learning-based baselines, demonstrating its superior performance on Arti4D. We make code and Arti4D publicly available at https://artipoint.cs.uni-freiburg.de.

Figures

Figures reproduced from arXiv: 2509.01708 by the authors.

Figure 1
Figure 1. We present ArtiPoint, our novel approach to articulated object estimation in the wild. ArtiPoint makes use of deep point tracking and factor graph optimization. It does not rely on costly scene-wise optimization or deep models that are prone to overfit. We evaluate our approach on the Arti4D dataset, a novel scene-level articulated object dataset that is recorded with a moving, ego-centric camera, compared to previo… view at source ↗
Figure 2
Figure 2. Overview of our method: We take an ego-centric RGB-D video as input and employ hand tracking [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results on Arti4D: Estimated joint axis and pose trajectory of a drawer in a kitchen scene [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 58 canonical work pages

  1. [1]

    M. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. Openvla: An open-source vision-language-action model. Conference on Robot Learning, 2024

    work page 2024

  2. [2]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. International Journal of Robotics Research, 2024

    work page 2024

  3. [3]

    Chisari, N

    E. Chisari, N. Heppert, M. Argus, T. Welschehold, T. Brox, and A. Valada. Learning robotic manipulation policies from point clouds with conditional flow matching. Conference on Robot Learning, 2024

    work page 2024

  4. [4]

    Gupta, M

    A. Gupta, M. Zhang, R. Sathua, and S. Gupta. Opening cabinets and drawers in the real world using a commodity mobile manipulator. Robotics: Science and Systems, 2025

    work page 2025

  5. [5]

    Nematollahi, E

    I. Nematollahi, E. Rosete-Beas, A. R¨ofer, T. Welschehold, A. Valada, and W. Burgard. Robot skill adaptation via soft actor-critic gaussian mixture models. In IEEE International Conference on Robotics and Automation, pages 8651–8657. IEEE, 2022

    work page 2022

  6. [6]

    Sturm, C

    J. Sturm, C. Plagemann, and W. Burgard. Unsupervised body scheme learning through self- perception. In IEEE International Conference on Robotics and Automation, pages 3328–3333, 2008

    work page 2008

  7. [7]

    Zhang, B

    H. Zhang, B. Eisner, and D. Held. Flowbot++: Learning generalized articulated objects manipu- lation via articulation projection. Conference on Robot Learning, 2023

    work page 2023

  8. [8]

    Honerkamp, T

    D. Honerkamp, T. Welschehold, and A. Valada. N2 m2: Learning navigation for arbitrary mobile manipulation motions in unseen and dynamic environments. IEEE Transactions on Robotics, 39 (5):3601–3619, 2023

    work page 2023

Show all 61 references
  1. [9]

    Buchanan, A

    R. Buchanan, A. R¨ofer, J. Moura, A. Valada, and S. Vijayakumar. Online estimation of articulated objects with factor graphs using vision and proprioceptive sensing. In IEEE International Conference on Robotics and Automation, pages 16111–16117, 2024

    2024

  2. [10]

    Schmalstieg, D

    F. Schmalstieg, D. Honerkamp, T. Welschehold, and A. Valada. Learning hierarchical interactive multi-object search for mobile manipulation. IEEE Robotics and Automation Letters , 8(12): 8549–8556, 2023

    2023

  3. [11]

    J. Liu, A. Mahdavi-Amiri, and M. Savva. Paris: Part-level reconstruction and motion analysis for articulated objects. In International Conference on Computer Vision, pages 352–363, 2023

    2023

  4. [12]

    Y . Liu, B. Jia, R. Lu, J. Ni, S.-C. Zhu, and S. Huang. Building interactable replicas of complex articulated objects via gaussian splatting. In International Conf. on Learning Representations, 2025

    2025

  5. [13]

    Honerkamp, M

    D. Honerkamp, M. B¨uchner, F. Despinoy, T. Welschehold, and A. Valada. Language-grounded dynamic scene graphs for interactive object search with mobile manipulation. IEEE Robotics and Automation Letters, 2024

    2024

  6. [14]

    Mohammadi, D

    M. Mohammadi, D. Honerkamp, M. B ¨uchner, M. Cassinelli, T. Welschehold, F. Despinoy, I. Gilitschenski, and A. Valada. More: Mobile manipulation rearrangement through grounded language reasoning. IEEE/RSJ International Conference on Intelligent Robots and Systems , 2025

    2025

  7. [15]

    Bandura and R

    A. Bandura and R. H. Walters. Social learning theory, volume 1. Prentice hall Englewood Cliffs, NJ, 1977

    1977

  8. [16]

    Doersch, Y

    C. Doersch, Y . Yang, M. Vecerik, D. Gokay, A. Gupta, Y . Aytar, J. Carreira, and A. Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In International Conference on Computer Vision, pages 10061–10072, 2023

    2023

  9. [17]

    Karaev, I

    N. Karaev, I. Rocco, B. Graham, N. Neverova, A. Vedaldi, and C. Rupprecht. Cotracker: It is better to track together. In European Conference on Computer Vision, pages 18–35, 2024

    2024

  10. [18]

    Sturm, C

    J. Sturm, C. Stachniss, and W. Burgard. A probabilistic framework for learning kinematic models of articulated objects, 2011

    2011

  11. [19]

    Pillai, M

    S. Pillai, M. Walter, and S. Teller. Learning articulated motions from visual demonstration. In Robotics: Science and Systems, Berkeley, USA, July 2014. doi:10.15607/RSS.2014.X.050

    2014 doi

  12. [20]

    D. Katz, A. Orthey, and O. Brock. Interactive perception of articulated objects. In International Symposium on Experimental Robotics, pages 301–315. Springer, 2014

    2014

  13. [21]

    H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, pages 404–417, 2006

    2006

  14. [22]

    Sturm, K

    J. Sturm, K. Konolige, C. Stachniss, and W. Burgard. 3d pose estimation, tracking and model learning of articulated objects from dense depth video using projected texture stereo. In RGB-D: Advanced Reasoning with Depth Cameras Workshop, RSS, 2010

    2010

  15. [23]

    A. Jain, R. Lioutikov, C. Chuck, and S. Niekum. Screwnet: Category-independent articulation model estimation from depth images using screw theory. In IEEE International Conference on Robotics and Automation, pages 13670–13677, 2021

    2021

  16. [24]

    A. Jain, S. Giguere, R. Lioutikov, and S. Niekum. Distributional depth-based estimation of object articulation models. In Conference on Robot Learning, pages 1611–1621, 2022

    2022

  17. [25]

    X. Li, H. Wang, L. Yi, L. J. Guibas, A. L. Abbott, and S. Song. Category-level articulated object pose estimation. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 3706–3715, 2020

    2020

  18. [26]

    Heppert, T

    N. Heppert, T. Migimatsu, B. Yi, C. Chen, and J. Bohg. Category-independent articulated object tracking with factor graphs. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3800–3807, 2022

    2022

  19. [27]

    V . Zeng, T. E. Lee, J. Liang, and O. Kroemer. Visual identification of articulated object parts. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2443–2450, 2021

    2021

  20. [28]

    Heppert, M

    N. Heppert, M. Z. Irshad, S. Zakharov, K. Liu, R. A. Ambrus, J. Bohg, A. Valada, and T. Kollar. Carto: Category and joint agnostic reconstruction of articulated objects. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 21201–21210, 2023

    2023

  21. [29]

    Huang, I

    X. Huang, I. Walker, and S. Birchfield. Occlusion-aware reconstruction and manipulation of 3d articulated objects. In IEEE International Conference on Robotics and Automation, pages 1365–1371, 2012

    2012

  22. [30]

    Mart´ın-Mart´ın, S

    R. Mart´ın-Mart´ın, S. H ¨ofer, and O. Brock. An integrated approach to visual perception of articulated objects. In IEEE International Conference on Robotics and Automation , pages 5091–5097. IEEE, 2016

    2016

  23. [31]

    Jiang, C.-C

    Z. Jiang, C.-C. Hsu, and Y . Zhu. Ditto: Building digital twins of articulated objects from interaction. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 5616–5626, 2022

    2022

  24. [32]

    D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91–110, 2004

    2004

  25. [33]

    Shi et al

    J. Shi et al. Good features to track. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 593–600, 1994

    1994

  26. [34]

    A. W. Harley, Z. Fang, and K. Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59–75, 2022

    2022

  27. [35]

    Doersch, A

    C. Doersch, A. Gupta, L. Markeeva, A. Recasens, L. Smaira, Y . Aytar, J. Carreira, A. Zisserman, and Y . Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022

    2022

  28. [36]

    Zheng, A

    Y . Zheng, A. W. Harley, B. Shen, G. Wetzstein, and L. J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In International Conference on Computer Vision, pages 19855–19865, 2023

    2023

  29. [37]

    S. Cho, J. Huang, J. Nam, H. An, S. Kim, and J.-Y . Lee. Local all-pair correspondence for point tracking. In European Conference on Computer Vision, pages 306–325, 2024

    2024

  30. [38]

    Doersch, P

    C. Doersch, P. Luc, Y . Yang, D. Gokay, S. Koppula, A. Gupta, J. Heyward, I. Rocco, R. Goroshin, J. Carreira, et al. Bootstap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, pages 3257–3274, 2024

    2024

  31. [39]

    Karaev, I

    N. Karaev, I. Makarov, J. Wang, N. Neverova, A. Vedaldi, and C. Rupprecht. Cotracker3: Sim- pler and better point tracking by pseudo-labelling real videos. arXiv preprint arXiv:2410.11831, 2024

    2024 arXiv

  32. [40]

    Zhang, D

    C. Zhang, D. Han, Y . Qiao, J. U. Kim, S.-H. Bae, S. Lee, and C. S. Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023

    2023 arXiv

  33. [41]

    Camporese

    G. Camporese. Hands segmentation is all you need. https://github.com/guglielmocamporese,

  34. [42]

    T. D. Barfoot. State estimation for robotics. Cambridge University Press, 2024

    2024

  35. [43]

    Dellaert and G

    F. Dellaert and G. Contributors. borglab/gtsam, May 2022. URL https://github.com/ borglab/gtsam)

    2022

  36. [44]

    Y . Weng, B. Wen, J. Tremblay, V . Blukis, D. Fox, L. Guibas, and S. Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3141–3150, 2024

    2024

  37. [45]

    Xiang, Y

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang, et al. Sapien: A simulated part-based interactive environment. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition

  38. [46]

    Mart´ın-Mart´ın, C

    R. Mart´ın-Mart´ın, C. Eppner, and O. Brock. The rbo dataset of articulated objects and interac- tions. International Journal of Robotics Research, 38(9):1013–1019, 2019

    2019

  39. [47]

    Rotondi, F

    D. Rotondi, F. Scaparro, H. Blum, and K. O. Arras. Fungraph: Functionality aware 3d scene graphs for language-prompted scene interaction. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025

    2025

  40. [48]

    Zhang, A

    C. Zhang, A. Delitzas, F. Wang, R. Zhang, X. Ji, M. Pollefeys, and F. Engelmann. Open- vocabulary functional 3d scene graphs for real-world indoor spaces. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2025

    2025

  41. [49]

    Kassab, S

    C. Kassab, S. Morin, M. B¨uchner, M. Mattamala, K. Gupta, A. Valada, L. Paull, and M. Fallon. Openlex3d: A new evaluation benchmark for open-vocabulary 3d scene representations. arXiv preprint arXiv:2503.19764, 2025

    2025

  42. [50]

    Delitzas, A

    A. Delitzas, A. Takmaz, F. Tombari, R. Sumner, M. Pollefeys, and F. Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  43. [51]

    Werby, C

    A. Werby, C. Huang, M. B¨uchner, A. Valada, and W. Burgard. Hierarchical Open-V ocabulary 3D Scene Graphs for Language-Grounded Robot Navigation. In Robotics: Science and Systems, Delft, Netherlands, July 2024. doi:10.15607/RSS.2024.XX.077

    2024 doi

  44. [52]

    Palazzolo, J

    E. Palazzolo, J. Behley, P. Lottes, P. Gigu`ere, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2019

    2019

  45. [53]

    C.-C. Hsu, Z. Jiang, and Y . Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In IEEE International Conference on Robotics and Automation, pages 3933–3939, 2023

    2023

  46. [54]

    Sturm, V

    J. Sturm, V . Pradeep, C. Stachniss, C. Plagemann, K. Konolige, and W. Burgard. Learning kine- matic models for articulated objects. In International Joint Conferences on Artificial Intelligence, 2009

    2009

  47. [55]

    Rublee, V

    E. Rublee, V . Rabaud, K. Konolige, and G. Bradski. Orb: An efficient alternative to sift or surf. In International Conference on Computer Vision, pages 2564–2571, 2011

    2011

  48. [56]

    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

    2021

  49. [57]

    Geneva, K

    P. Geneva, K. Eckenhoff, W. Lee, Y . Yang, and G. Huang. OpenVINS: A research platform for visual-inertial estimation. In IEEE International Conference on Robotics and Automation, 2020

    2020

  50. [58]

    Murai, E

    R. Murai, E. Dexheimer, and A. J. Davison. MASt3R-SLAM: Real-time dense SLAM with 3D reconstruction priors. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2025

    2025

  51. [59]

    Teed and J

    Z. Teed and J. Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in Neural Information Processing Systems, 2021

    2021

  52. [60]

    H. Lim, D. Kim, G. Shin, J. Shi, I. Vizzo, H. Myung, J. Park, and L. Carlone. KISS-Matcher: Fast and Robust Point Cloud Registration Revisited. In IEEE International Conference on Robotics and Automation, 2025. Articulated Object Estimation in the Wild: Supplementary Material ...

    2025

  53. [2021]

    URL https://github.com/guglielmocamporese/ hands-segmentation-pytorch

Pith tools

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