Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

SplArt: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SplArt claims a two-part articulated object can be digitized from two sets of posed RGB photos, with its hinge or slider joint estimated and no 3D supervision, and re-rendered in real time.

desk verdict Solid new self-supervised 3DGS method for two-part articulated objects from RGB, but the ablation contradicts the robustness story and the real-world SfM dependency is unvalidated. read the letter →

arxiv 2506.03594 v1 pith:DUJOYQBV submitted 2025-06-04 cs.GR cs.CVcs.LGcs.MMcs.RO

Shengjie Lin , Jiading Fang , Muhammad Zubair Irshad , Vitor Campagnolo Guizilini , Rares Andrei Ambrus , Greg Shakhnarovich , Matthew R. Walter This is my paper
classification cs.GRcs.CVcs.LGcs.MMcs.RO
keywords 3DGaussianSplattingarticulatedobjectreconstructionarticulationestimationpart-levelsegmentationself-supervisedlearningnovelviewsynthesiskinematicmodel
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

SplArt claims that a two-part articulated object — a static base plus a part that swings on a hinge or slides along a rail — can be fully digitized from two batches of posed RGB photos, one batch per articulation state, with no 3D supervision, no part annotations, and no category-specific training. If true, an ordinary handheld camera is enough to produce a photorealistic, re-animatable digital twin of everyday objects such as drawers, scissors, monitors, and cabinets, with the joint axis, pivot, and range of motion recovered automatically. The method gives every 3D Gaussian a learnable mobility value that splits it into static and mobile components, then uses geometric self-supervision (weighted Chamfer distances between the two states' reconstructions) followed by photometric refinement in three decoupled stages. On the established PARIS benchmark the paper reports 99% success with joint-axis error roughly an order of magnitude below the PARIS baseline, and 85.5% on its own harder 22-object benchmark. Rendering new views and new articulation states runs at 60–100 frames per second, a speed regime NeRF-based alternatives do not reach.

What carries the argument

The load-bearing object is the mobility-augmented Gaussian: each Gaussian carries a scalar $m \in [0,1]$ that decomposes it into a static component of opacity $\sigma(1-m)$ and a mobile component of opacity $\sigma m$, both inheriting the original position, rotation, and scale. Two representation identities carry the optimization. The cross-static representation $$G_s = $G^{{\mathrm{tgt}}$}_l \cdot (1 - M_l) \;\oplus\; $G^{{\mathrm{tgt}}$}_{1-l} \cdot (1 - M_{1-l})$$ asserts that the static parts in the two states are the same geometry, so it is the supervision signal for mobility estimation. The cross-mobile representation $$G^m_l = $G^{{\mathrm{tgt}}$}_l \cdot M_l \;\oplus\; {}_l T_{1-l}\left($G^{{\mathrm{tgt}}$}_{1-l} \cdot M_{1-l}\right)$$ asserts that the two states' mobile parts are related by the articulation transform, so it is the signal for articulation estimation. Both are evaluated by a weighted Chamfer distance whose weights are the mobility-modified opacities, giving geometric supervision that operates without rendering. The articulation transform updates each Gaussian by $\mu' = R_{a,\theta}(\mu-p)+p+da$ and $R' = R_{a,\theta}R$ while leaving scale and opacity fixed, and it compensates the radiance-field query direction by $R_{a,\theta}^{-1}d$. The three stages — per-state reconstruction, mobility estimation (geometric then photometric), then articulation estimation with randomized restarts over two geometric formulations followed by cross-mobile rendering — sequence these pieces so that each stage hands the next a trustworthy initialization.

What would settle it

Capture two states of a hinged object separated by only a small fraction of the joint's range (say 2–3 degrees of rotation), run the pipeline, and compare the estimated axis to the ground-truth hinge line: at near-zero spanned angles the cross-mobile and cross-static Chamfer alignments become nearly indifferent to axis direction, so a still-accurate axis would show the geometric self-supervision is stronger than the two-end-state design suggests, while a large error would confirm the span is load-bearing. A second decisive test runs the two-state protocol on a three-part object and checks whether the single-joint model silently fits a spurious axis instead of flagging the violation that the conclusion concedes.

Watch

Extended reading notes

Core claim

The central claim is that articulation estimation and part-level reconstruction reduce to a single learnable quantity: a mobility value $m \in [0,1]$ carried by each Gaussian in a 3D Gaussian Splatting scene representation, where the scene is a cloud of Gaussian blobs with position, rotation, scale, opacity, and color. A Gaussian with mobility $m$ acts as a static component of opacity $\sigma(1-m)$ plus a mobile component of opacity $\sigma m$, and the two observed articulation states are stitched together in two complementary ways: the cross-static representation concatenates the static components of both states with the mobile component of the target state, while the cross-mobile representation further pulls in the other state's mobile component transformed by the unknown motion $x' = R_{a,\theta}(x-p)+p+da$. A weighted Chamfer distance using the mobility-modified opacities as point weights supplies geometric self-supervision that needs no images at all, and cross-static and cross-mobile rendering apply ordinary photometric loss. The three-stage schedule — independent per-state reconstruction, then mobility estimation, then joint refinement of articulation, mobility, and Gaussians — is the safeguard the paper identifies against the local optima that plague end-to-end joint optimization. Positioned as the first self-supervised 3DGS method for this task, the paper reports 99% success on PARIS-PMS and 85.5% on its own harder SPLART-PMS, with joint-axis error roughly an order of magnitude below the PARIS baseline and part-level segmentation produced as a byproduct.

Load-bearing premise

The load-bearing premise is that the object is exactly two rigid parts joined by a single revolute or prismatic joint whose type is known before optimization (Section 3.1 states the type is assumed), that the two photographed states span the motion well enough for geometric consistency to disambiguate the axis, and that the real-world preprocessing — segmentation and structure-from-motion — yields a consistent object foreground in one shared world frame; if any of these fails, the estimated articulation no longer targets the true motion, and the method offers no mechanism for multi-joint objects.

Editorial extensions

If this is right

  • Two sets of posed RGB photos at two articulation states are sufficient input; no depth sensor, 3D supervision, part annotation, or category-specific model is required.
  • The estimated articulation is directly executable: the axis, pivot, and angle (or translation axis and distance) define the motion a robot can reproduce or an AR system can animate.
  • Novel-view and novel-articulation rendering is photorealistic and real-time (60–100 fps), a regime NeRF-based baselines do not reach.
  • Part segmentation is a byproduct of the optimized mobility values, yielding static/mobile masks (mean IoU 0.957 on PARIS-PMS) without extra labels.
  • The new SPLART-PMS benchmark samples articulation states from $[-0.1, 1.1]$ rather than binary endpoints, so future methods can be scored quantitatively on novel-articulation synthesis.

Reading between the lines

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

  • Editorial inference: the per-Gaussian mobility trick is a general recipe — one scalar per primitive gating its participation in a shared rigid transform — and could be stacked as one mobility field per joint to cover multi-part objects, a direction the paper names as future work.
  • Editorial inference: because the geometric stages need no rendering, the method's true input requirement is not 'two states' but 'two states spanning enough of the joint's range'; shrinking the spanned angle or translation is the natural stress test of this dependence.
  • Editorial inference: the known-articulation-type assumption could be lifted by fitting revolute and prismatic models separately and selecting the one with lower cross-mobile loss, which the paper's multi-try scheme already resembles.
  • Editorial inference: the failure modes shown with a clock hand rotating on an unexpected axis and a globe spinning alone reveal that the optimizer finds a moving explanation rather than the intended one; photographing three or more states instead of two would break most such ambiguities.
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

2 major / 5 minor

Summary. The paper introduces SPLART, a self-supervised 3D Gaussian Splatting (3DGS) method for reconstructing a two-part articulated object and estimating its revolute or prismatic joint from two sets of posed RGB images captured at two articulation states. The key ideas are (i) a differentiable mobility parameter per Gaussian that splits each Gaussian into static and mobile components, and (ii) a three-stage optimization pipeline that decouples per-state reconstruction, mobility estimation via cross-static geometric and photometric consistency, and articulation estimation via cross-mobile geometric and photometric consistency. Evaluations are reported on the PARIS PartNet-Mobility subset, a new SPLART-PMS dataset, and a qualitative real-world dataset; the reported success rates and articulation errors are very favorable relative to the PARIS baseline, and the method is compared with DTA, which requires ground-truth depth. The code is publicly available.

Significance. If the results are reproducible, this is a valuable contribution: it is the first self-supervised 3DGS-based articulated object reconstruction method that works without 3D supervision or category-specific priors, and the real-time rendering advantage of 3DGS is a clear practical benefit. The Gaussian articulation transform in the supplementary material (Eqs. 18-21) is a clean and correctly derived result. The paper also follows several good evaluation practices: randomized trials with ten seeds, per-scene results, success-rate reporting, and a dedicated analysis of the geometric-consistency optimization landscape (Suppl. Fig. 8). However, the strength of the claims is currently undermined by an unexplained counterexample in the ablation study and by the lack of quantitative validation of the real-world pipeline.

major comments (2)
  1. [Section 4.4, Table 4] In the revolute rows of Table 4, the full SPLART method reports err_a = 20.76±0.81 (×10⁻² DEG) with a 77.5% success rate, whereas the SPLART-2a ablation (which removes Stage 2(a)) reports err_a = 2.91±0.55 with a 78.5% success rate; the full method is also worse on err_p and err_r. This result contradicts the paper's claim (abstract, Section 1, and Section 4.4) that the multi-stage optimization significantly improves accuracy and robustness. The paper does not discuss this counterintuitive outcome. The authors should either provide a mechanism that explains the degradation (e.g., additional local minima introduced by Stage 2(a), or differences in the subset of runs over which the averages are computed) or present the full error distributions across all ten runs per scene. As it stands, the ablation evidence does not support the central claim that the complete pipeline is more accurate than its components.
  2. [Section 3.3, Table 3] The real-world practicality claim rests on the preprocessing pipeline that uses SAM 2 and COLMAP to place both articulation states into a common world coordinate frame. All quantitative experiments in Section 4 use synthetic data with perfect shared poses, so they do not exercise this assumption. The real-world results (Table 3) are qualitative only; no information is given about COLMAP alignment errors, the fraction of objects for which the pipeline succeeded, or a comparison of the estimated articulation parameters against any ground truth. Since the geometric consistency losses (Eqs. 6, 10, 11) and the articulation optimization (Eq. 10) are meaningful only if both state reconstructions share a single world frame, an unquantified sensitivity to cross-state pose alignment is a significant gap. Please provide a quantitative real-world evaluation (e.g., pose accuracy from COLMAP, or articulation errors measured against manual annotations) or scale back the 'real-world practicality' statement.
minor comments (5)
  1. [Section 3.1 and Abstract] The paper states in Section 3.1 that the articulation type (revolute or prismatic) is assumed known, but the abstract and introduction do not mention this restriction; please state it upfront, since the user must specify the joint type.
  2. [Equation (7)] The minimization over M_{1-l} is unclear as written because the loss also depends on M_l through G^{tgt,cs}_l; please clarify the symmetric optimization over both states (l=0 and l=1) or write the total objective.
  3. [Tables 1 and 4 and supplementary tables] The column headers for axis error and other metrics have an unusual formatting (the 'x̄ err_a' prefix appears to be malformed); please clean up the table headers for readability.
  4. [Section 4.2, success criteria] The success criteria use thresholds err_a < 5 DEG and err_r < 10 DEG, which are much looser than the typical errors achieved by the method (often below 1 DEG); consider also reporting a stricter success criterion to better separate methods.
  5. [Section 4.1, dataset availability] The new SPLART-PMS dataset is not released (only code is linked); for reproducibility, please provide a download link or state an intention to release the dataset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found: articulation and mobility are estimated by optimizing photometric and geometric losses against the input observations, with external benchmark ground truth used for evaluation.

full rationale

SplArt's derivation chain is self-contained. Stage 1 builds a separate 3DGS reconstruction for each articulation state directly from the posed RGB images via the photometric loss (Eqs. 2a/2b). Stage 2 estimates per-Gaussian mobilities by minimizing weighted Chamfer distances to those reference reconstructions (Eqs. 5, 6a, 6b) and by cross-static rendering against the input images (Eqs. 7a/7b). Stage 3 estimates the articulation transform by matching the mobile components across states through Chamfer losses (Eqs. 10-13) and refines it with photometric supervision (Eqs. 14a/14b). At no point is a target quantity defined in terms of the quantity being predicted, nor is any fitted parameter renamed as a prediction: the articulation parameters and mobilities are free variables optimized from the same observations used to produce the reconstructions, and the paper's reported success rates and errors are computed against ground-truth articulation parameters, depth, and part segmentation from PARIS-PMS and SPLART-PMS. The self-referential geometric consistency losses are a legitimate self-supervision signal, not a circular definition, because the reference reconstructions themselves are derived from the input images. The only self-citations are to related prior work (CARTO, ShAPO, ReFiNe, etc.) and are not load-bearing for the method's correctness. The unquantified dependence on COLMAP cross-state alignment in the real-world pipeline is a validation gap, but it is not circular reasoning.

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

The method relies on a two-rigid-part, known-joint-type model and on 3DGS as its backbone. The mobility parameter is a model-internal attribute, not an externally visible entity. Several hyperparameters that control the losses are not given numerical values, which is a reproducibility gap.

free parameters (4)
  • mobility regularization weight lambda_geom_m = not reported in paper
    Appears in Eq. 6a as lambda_geom_m; chosen by hand, value not given, affects the mobility segmentation bias toward static parts.
  • mobility regularization weight lambda_photo_m = not reported in paper
    Appears in Eq. 7a as lambda_photo_m; chosen by hand, value not given.
  • randomized tries K_m, K_cm = 3
    From supplementary Sec. 8; controls how many random restarts Stage 3(a) uses and influences robustness.
  • stage iteration counts = 10k, 5k, 10k
    From supplementary Sec. 8; chosen by hand, fixed across experiments.
assumptions (5)
  • domain assumption The object is composed of exactly two rigid parts: a static parent and a child part moving via a single revolute or prismatic joint.
    Section 3.1 formalizes the problem as two rigid parts and assumes a known articulation type; this is the core modeling premise.
  • domain assumption The articulation type (revolute or prismatic) is known a priori.
    Section 3.1: "Assuming a known articulation type".
  • domain assumption Two sets of posed RGB images at two end states provide sufficient information for reconstruction and articulation estimation.
    Section 3.1 defines input as two state observations; if the two states do not disambiguate the motion (e.g., only one end visible), the method can fail, as the paper acknowledges for the pen scene.
  • standard math 3D Gaussian Splatting with opacity-weighted alpha compositing is a valid renderer for photometric self-supervision.
    The method builds on the published 3DGS formulation (Ref. 26); the photometric loss in Eq. 2a assumes this renderer is differentiable and accurate.
  • domain assumption For real-world scenes, SAM 2 segmentation and COLMAP/SuperPoint/SuperGlue produce a consistent foreground mask and a shared world coordinate system across both states.
    Section 3.3 describes this pipeline; errors in segmentation or pose estimation propagate into the reconstruction.
invented entities (1)
  • Per-Gaussian mobility parameter m
    purpose: Splits each Gaussian into static and mobile components via opacity scaling for part segmentation.
    This is a new model attribute introduced by the paper, optimized by geometric and photometric losses; it has no falsifiable handle outside the method itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SplArt: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/DUJOYQBV

@misc{pith2026250603594,
  author       = {Pith},
  title        = {Pith review of: SplArt: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUJOYQBV}},
  note         = {Machine review of arXiv:2506.03594}
}
read the original abstract

Reconstructing articulated objects prevalent in daily environments is crucial for applications in augmented/virtual reality and robotics. However, existing methods face scalability limitations (requiring 3D supervision or costly annotations), robustness issues (being susceptible to local optima), and rendering shortcomings (lacking speed or photorealism). We introduce SplArt, a self-supervised, category-agnostic framework that leverages 3D Gaussian Splatting (3DGS) to reconstruct articulated objects and infer kinematics from two sets of posed RGB images captured at different articulation states, enabling real-time photorealistic rendering for novel viewpoints and articulations. SplArt augments 3DGS with a differentiable mobility parameter per Gaussian, achieving refined part segmentation. A multi-stage optimization strategy is employed to progressively handle reconstruction, part segmentation, and articulation estimation, significantly enhancing robustness and accuracy. SplArt exploits geometric self-supervision, effectively addressing challenging scenarios without requiring 3D annotations or category-specific priors. Evaluations on established and newly proposed benchmarks, along with applications to real-world scenarios using a handheld RGB camera, demonstrate SplArt's state-of-the-art performance and real-world practicality. Code is publicly available at https://github.com/ripl/splart.

Figures

Figures reproduced from arXiv: 2506.03594 by the authors.

Figure 1
Figure 1. Given (top) RGB images of an object at two articulation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Methodology overview of SPLART. It consists of three decoupled stages for the purpose of stable optimization. Stage 1 constructs a 3DGS for each state from posed RGB images using the photometric loss. Stage 2 proposes cross-static formulation for mobility estimation. Intuitively, it combines static parts of both states and the mobile part of the desired state for the target Gaussians. Stage 3 proposes cross-mobile f… view at source ↗
Figure 3
Figure 3. 103549-Toaster has a flat slider as its mobile part, sim [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: State interpolation from the estimated articulation. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Failure cases due to inherent ambiguities in articulation. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Since the radiance field is not modified by the articula [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: An analysis of the robustness of geometric consistency for articulation estimation. The horizontal axis represents different [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Additional visualizations of state interpolation for three objects along with their estimated articulation models. The left- and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Additional qualitative comparisons of part-level reconstruction and articulation estimation. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

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. Full citation record

  1. StructureGS: Structure-aware Gaussian Splatting for Articulated Object Reconstruction

    cs.GR 2026-07 conditional novelty 6.0 of 10

    OBB-based part-fitting and contact losses on 3D Gaussians disentangle geometry, appearance, and motion for cleaner articulated reconstruction than photometric-only baselines.

Reference graph

Works this paper leans on

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

  1. [1]

    Learning to generalize kinematic models to novel objects

    Ben Abbatematteo, Stefanie Tellex, and George Konidaris. Learning to generalize kinematic models to novel objects. In Proceedings of the Conference on Robot Learning (CoRL),

    work page

  2. [2]

    ShapeNet: An information-rich 3D model repository.arXiv preprint arXiv:1512.03012, 2015

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository.arXiv preprint arXiv:1512.03012, 2015. 5

    arXiv 2015

  3. [3]

    3D-R2N2: A unified approach for single and multi-view 3d object reconstruction

    Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. InPro- ceedings of the European Conference on Computer Vision (ECCV), 2016. 2

    work page 2016

  4. [4]

    A volumetric method for building complex models from range images

    Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. InProceeding of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 303–312, 1996. 7

    work page 1996

  5. [5]

    Daniele, Thomas M

    Andrea F. Daniele, Thomas M. Howard, and Matthew R. Walter. A multiview approach to learning articulated mo- tion models. InProceedings of the International Symposium of Robotics Research (ISRR), 2017. 1

    work page 2017

  6. [6]

    Articulate your nerf: Unsupervised articulated object modeling via con- ditional view synthesis.Advances in Neural Information Processing Systems, 37:119717–119741, 2025

    Jianning Deng, Kartic Subr, and Hakan Bilen. Articulate your nerf: Unsupervised articulated object modeling via con- ditional view synthesis.Advances in Neural Information Processing Systems, 37:119717–119741, 2025. 1, 2, 3, 7, 9

    work page 2025

  7. [7]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabi- novich. Superpoint: Self-supervised interest point detection and description. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 224–236, 2018. 5

    work page 2018

  8. [8]

    A point set generation network for 3D object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3D object reconstruction from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 605–613, 2017. 2

    work page 2017

Show all 72 references
  1. [9]

    Act the part: Learning interaction strategies for articulated object part discovery

    Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. InProceedings of the International Confer- ence on Computer Vision (ICCV), pages 15752–15761, 2021. 2

    2021

  2. [10]

    PartManip: Learning cross-category generalizable part manipulation policy from point cloud ob- servations

    Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. PartManip: Learning cross-category generalizable part manipulation policy from point cloud ob- servations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2...

    2023

  3. [11]

    GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts

    Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  4. [12]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering.CVPR, 2024

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering.CVPR, 2024. 2

    2024

  5. [13]

    CARTO: Category and joint agnostic reconstruction of articulated objects

    Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Rares Andrei Ambrus, Jeannette Bohg, Ab- hinav Valada, and Thomas Kollar. CARTO: Category and joint agnostic reconstruction of articulated objects. InPro- ceedings of the IEEE/CVF Conference on Computer Visi...

    2023

  6. [14]

    Ditto in the house: Building articulation models of indoor scenes through interactive perception

    Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3933–3939, 2023. 2

    2023

  7. [15]

    Learning to predict part mo- bility from a single static snapshot.ACM Transactions On Graphics (TOG), 36(6):1–13, 2017

    Ruizhen Hu, Wenchao Li, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Learning to predict part mo- bility from a single static snapshot.ACM Transactions On Graphics (TOG), 36(6):1–13, 2017. 2

    2017

  8. [16]

    Occlusion- aware reconstruction and manipulation of 3d articulated ob- jects

    Xiaoxia Huang, Ian Walker, and Stan Birchfield. Occlusion- aware reconstruction and manipulation of 3d articulated ob- jects. InProceedings of the IEEE International Confer- ence on Robotics and Automation (ICRA), pages 1365–1371,

  9. [17]

    CenterSnap: Single-shot multi- object 3D shape reconstruction and categorical 6D pose and size estimation

    Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. CenterSnap: Single-shot multi- object 3D shape reconstruction and categorical 6D pose and size estimation. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2022. 2

    2022

  10. [18]

    ShAPO: Im- plicit representations for multi-object shape, appearance, and pose optimization

    Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Im- plicit representations for multi-object shape, appearance, and pose optimization. InProceedings of the European Confer- ence on Computer Vision (ECCV), pages 275–292...

    2022

  11. [19]

    Neural fields in robotics: A survey

    Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv:2410.20220, 2024. 2

    2024 arXiv

  12. [20]

    ScrewNet: Category-independent articulation model estimation from depth images using screw the- ory

    Ajinkya Jain, Rudolf Lioutikov, Caleb Chuck, and Scott Niekum. ScrewNet: Category-independent articulation model estimation from depth images using screw the- ory. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 13670–13677,

  13. [21]

    Prentice-Hall, Inc., 1989

    Anil K Jain.Fundamentals of digital image processing. Prentice-Hall, Inc., 1989. 2

    1989

  14. [22]

    OPD: Single-view 3D openable part detection

    Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X Chang. OPD: Single-view 3D openable part detection. In European Conference on Computer Vision, pages 410–426. Springer, 2022. 2

    2022

  15. [23]

    Ditto: Building digital twins of articulated objects from interaction

    Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5616–5626,

  16. [24]

    Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects

    Dov Katz, Moslem Kazemi, J Andrew Bagnell, and Anthony Stentz. Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects. InProceed- ings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5003–5010, 2013. 1

    2013

  17. [25]

    Unsu- pervised pose-aware part decomposition for man-made artic- ulated objects

    Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Unsu- pervised pose-aware part decomposition for man-made artic- ulated objects. InProceedings of the European Conference on Computer Vision (ECCV), pages 558–575, 2022. 2

    2022

  18. [26]

    3D gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):139–1, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):139–1, 2023. 1, 2

    2023

  19. [27]

    Robot see robot do: Part-centric feature fields for visual imitation of ar- ticulated objects

    Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qian- qian Wang, Angjoo Kanazawa, and Ken Goldberg. Robot see robot do: Part-centric feature fields for visual imitation of ar- ticulated objects. InProceedings of the Conference on Robot Learning (CoRL), 2024. 2

    2024

  20. [28]

    Articulate-Anything: Auto- matic modeling of articulated objects via a vision-language foundation model.arXiv preprint arXiv:2410.13882, 2024

    Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Yecheng Jason Ma, Kyle Vedder, Arjun Krishna, Di- nesh Jayaraman, and Eric Eaton. Articulate-Anything: Auto- matic modeling of articulated objects via a vision-language foundation model.arXiv preprint arXiv:2410.13882, 2024. 2

    arXiv

  21. [29]

    Category-level articulated object pose estimation

    Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 3706–3715, 2020. 1, 2

    2020

  22. [30]

    PARIS: Part-level reconstruction and motion analysis for articulated objects

    Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. PARIS: Part-level reconstruction and motion analysis for articulated objects. InProceedings of the International Conference on Computer Vision (ICCV), pages 352–363, 2023. 1, 2, 3, 5, 7

    2023

  23. [31]

    AKB-48: A real- world articulated object knowledge base

    Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao- jun Yu, Yang Han, and Cewu Lu. AKB-48: A real- world articulated object knowledge base. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809–14818, 2022. 2

    2022

  24. [33]

    Build- ing rearticulable models for arbitrary 3D objects from 4D point clouds

    Shaowei Liu, Saurabh Gupta, and Shenlong Wang. Build- ing rearticulable models for arbitrary 3D objects from 4D point clouds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21138–21147, 2023. 2

    2023

  25. [34]

    Building interactable replicas of complex articulated objects via Gaussian splatting

    Yu Liu, Baoxiong Jia, Ruijie Lu, Junfeng Ni, Song-Chun Zhu, and Siyuan Huang. Building interactable replicas of complex articulated objects via Gaussian splatting. InPro- ceedings of the International Conference on Learning Rep- resentations (ICLR), 2025. 1, 2, 9

    2025

  26. [35]

    Marching cubes: A high resolution 3D surface construction algorithm

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. InSemi- nal graphics: pioneering efforts that shaped the field, pages 347–353. 1998. 7

    1998

  27. [36]

    Real2Code: Reconstruct articulated objects via code genera- tion.arXiv preprint arXiv:2406.08474, 2024

    Zhao Mandi, Yijia Weng, Dominik Bauer, and Shuran Song. Real2Code: Reconstruct articulated objects via code genera- tion.arXiv preprint arXiv:2406.08474, 2024. 2

    2024 arXiv

  28. [37]

    Pose estimation for augmented reality: a hands-on survey

    Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey. IEEE Transactions on Visualization and Computer Graph- ics, 22(12):2633–2651, 2015. 1

    2015

  29. [38]

    Occupancy networks: Learning 3D reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460–4470, 2019. 2

    2019

  30. [39]

    NeRF: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2, 1

    2021

  31. [40]

    Chang, Li Yi, Subarna Tripathi, Leonidas J

    Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large- scale benchmark for fine-grained and hierarchical part-level 3D object understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2019

  32. [41]

    A-SDF: Learning disentangled signed distance functions for articulated shape 10 representation

    Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, and Xiaolong Wang. A-SDF: Learning disentangled signed distance functions for articulated shape 10 representation. InProceedings of the International Confer- ence on Computer Vision (ICCV), pages 13001–13...

    2021

  33. [42]

    KinectFusion: Real-time dense surface mapping and track- ing

    Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. KinectFusion: Real-time dense surface mapping and track- ing. InProceedings of the IEEE International Symposium o...

  34. [43]

    Structure from action: Learning interactions for articulated object 3d structure discovery.arXiv preprint arXiv:2207.08997, 2022

    Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shu- ran Song. Structure from action: Learning interactions for articulated object 3d structure discovery.arXiv preprint arXiv:2207.08997, 2022. 2

    2022 arXiv

  35. [44]

    Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction

    Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589–5599, 2021. 2

    2021

  36. [45]

    DeepSDF: Learning continuous signed distance functions for shape representa- tion

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representa- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 165– 1...

    2019

  37. [46]

    Walter, and Seth Teller

    Sudeep Pillai, Matthew R. Walter, and Seth Teller. Learning articulated motions from visual demonstrations. InProceed- ings of Robotics: Science and Systems (RSS), 2014. 1

    2014

  38. [47]

    SAM 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024. 5

    2024 arXiv

  39. [48]

    SuperGlue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938–4947, 2020. 5

    2020

  40. [49]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104–4113, 2016. 5

    2016

  41. [50]

    REACTO: Reconstructing articulated objects from a single video

    Chaoyue Song, Jiacheng Wei, Chuan Sheng Foo, Guosheng Lin, and Fayao Liu. REACTO: Reconstructing articulated objects from a single video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5384–5395, 2024. 2

    2024

  42. [51]

    A probabilistic framework for learning kinematic models of ar- ticulated objects.Journal of Artificial Intelligence Research, 41:477–526, 2011

    J ¨urgen Sturm, Cyrill Stachniss, and Wolfram Burgard. A probabilistic framework for learning kinematic models of ar- ticulated objects.Journal of Artificial Intelligence Research, 41:477–526, 2011. 1

    2011

  43. [52]

    Augmented reality and robotics: A sur- vey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces

    Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A sur- vey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. InProceedings of the CHI Conference on Human Factors in Computing Systems, 2022. 1

    2022

  44. [53]

    Neural geometric level of detail: Real-time rendering with implicit 3D shapes

    Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer...

    2021

  45. [54]

    Nerfstudio: A mod- ular framework for neural radiance field development

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

    2023

  46. [55]

    CLA-NeRF: Category-level articulated neural radi- ance field

    Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. CLA-NeRF: Category-level articulated neural radi- ance field. InProceedings of the IEEE International Confer- ence on Robotics and Automation (ICRA), pages 8454–8460,

  47. [56]

    NeuS: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction.arXiv preprint arXiv:2106.10689, 2021

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction.arXiv preprint arXiv:2106.10689, 2021. 2

    2021 arXiv

  48. [57]

    Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes

    Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qin- ping Zhao, and Kai Xu. Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 8876–8884, 2019. 1, 2

    2019

  49. [58]

    A system- atic review of XR-based remote human-robot interaction sys- tems.arXiv preprint arXiv:2403.11384, 2024

    Xian Wang, Luyao Shen, and Lik-Hang Lee. A system- atic review of XR-based remote human-robot interaction sys- tems.arXiv preprint arXiv:2403.11384, 2024. 1

    2024 arXiv

  50. [59]

    Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 2

    2004

  51. [60]

    Self-supervised neural articulated shape and appearance models

    Fangyin Wei, Rohan Chabra, Lingni Ma, Christoph Lassner, Michael Zollh ¨ofer, Szymon Rusinkiewicz, Chris Sweeney, Richard Newcombe, and Mira Slavcheva. Self-supervised neural articulated shape and appearance models. InProceed- ings of the IEEE/CVF Conference on Computer Vision...

    2022

  52. [61]

    Neural implicit representation for building digital twins of unknown articulated objects

    Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. InProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 1...

    2024

  53. [62]

    Chang, Leonidas J

    Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive envi- ronment. InProceedings of the IEEE/CVF Conference on C...

    2020

  54. [63]

    RPM-Net: Recurrent prediction of motion and parts from point cloud

    Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver Van Kaick, Hao Zhang, and Hui Huang. RPM-Net: Recurrent prediction of motion and parts from point cloud. arXiv preprint arXiv:2006.14865, 2020. 2

    2006 arXiv

  55. [64]

    gsplat: An open-source library for Gaussian splatting, 2024

    Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting, 2024. 1 11

    2024

  56. [65]

    Deep part induction from articu- lated object pairs.arXiv preprint arXiv:1809.07417, 2018

    Li Yi, Haibin Huang, Difan Liu, Evangelos Kalogerakis, Hao Su, and Leonidas Guibas. Deep part induction from articu- lated object pairs.arXiv preprint arXiv:1809.07417, 2018. 2

    2018 arXiv

  57. [66]

    ReFiNe: Recursive field networks for cross-modal multi-scene representation

    Sergey Zakharov, Katherine Liu, Adrien Gaidon, and Rares Ambrus. ReFiNe: Recursive field networks for cross-modal multi-scene representation. InProceeding of the Interna- tional Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 2

    2024

  58. [67]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 2 12 SPLART:...

    2018

  59. [68]

    3D Gaussian Splatting 3D Gaussian Splatting [26] (3DGS) is a method for re- constructing 3D scenes from posed images by represent- ing the scene using Gaussian distributions in a continu- ous 3D space. Given a Gaussian blob parameterized by (µ, R, S, σ), whereµ∈R3 denotes the ...

  60. [69]

    We then need to update the parameters of the Gaussian (µ′, R′, S′, σ′)to reflect this articulation

    Articulating the Gaussians In SPLART, we articulate a Gaussian blob by (1) rotat- ing by an angleθaround a line specified by(p,a), where pdenotes the pivot point andadenotes the axis direc- tion; and (2) translating along the same line by distance d. We then need to update the...

  61. [70]

    We apply consistent hyper- parameters across all experiments, spanning synthetic and real-world datasets

    Implementation SPLARTleveragesnerfstudio[54] andgsplat[64], widely used open-source libraries for neural rendering and Gaus- 1 sian splatting, respectively. We apply consistent hyper- parameters across all experiments, spanning synthetic and real-world datasets. Optimization s...

  62. [71]

    Robustness of Geometric Consistency In Stage 3(a), we propose a practical strategy that involves multiple attempts using both mobile-only and cross-mobile geometric consistency for robust articulation estimation. For this approach to be effective, two prerequisites must be met...

  63. [72]

    Additional Qualitative Results We present more qualitative results, including comparison with PARIS [30] on part-level reconstruction and articula- tion estimation in Figure 10, and view synthesis for interpo- lated articulation states in Figure 9

  64. [73]

    Specifically, see Tab

    Per-Scene Quantitative Results We report quantitative comparison of SPLARTand PARIS for each scene in both PARIS-PMS and SPLART-PMS. Specifically, see Tab. 7, Tab. 9, and Tab. 10 for articula- tion estimation results on PARIS-PMS, revolute cases of SPLART-PMS, and prismatic ca...

Pith tools

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