Pith. sign in

REVIEW 4 major objections 6 minor 60 references

An articulated object's movable parts and joints can be recovered from a single closed configuration, with no observed motion, by cross-checking noisy model predictions against a shared mesh.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Rest2Art reconstructs part geometry and joint parameters of openable objects from a single closed-state observation by co-refining vision-language and segmentation outputs on a mesh and fitting joint models to video-diffusion-generated motion hypotheses.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A well-engineered rest-state reconstruction pipeline with a genuinely new VLM/SAM co-refinement plus VDM hypothesis-fitting loop, but the abstract overstates novelty and the evaluation overreaches on baseline fairness and metric-scale scope. the 4 major comments →

arxiv 2607.27749 v1 pith:N2CGE635 submitted 2026-07-30 cs.CV cs.RO

Articulated Object Reconstruction from Rest-State Observation

Daeun Lee , Jaeah Lee , Woosung Kim , Haebeom Jung , Jaesik Park This is my paper
classification cs.CV cs.RO
keywords articulated objectsrest-state reconstructionpart segmentationjoint estimationmesh representationvideo diffusiondigital twinkinematic structure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that articulated objects can be reconstructed from a single closed (rest-state) observation alone, something previously thought to require seeing the object in motion. The key move is to treat an explicit 3D mesh as the anchor of truth: a vision-language model and a segmentation model are iteratively co-refined so their disagreements correct each other, and a video diffusion model generates hypothetical articulation motions that are only used as weak hypotheses. The final joint parameters are then fitted geometrically against the mesh, which resolves monocular depth ambiguity and penalizes physically impossible inter-part penetration. If correct, this would let digital twins of furniture and other openable objects be built from ordinary closed-state photos or scans, without motion capture or part-count priors.

Core claim

Rest2Art claims that an articulated object's parts and joints can be recovered from a single closed-configuration mesh, with no observed motion and no prior knowledge of the number of parts. It argues that the missing motion information can be supplied by pretrained semantic and motion priors, provided their noisy outputs are reconciled against one another and against an explicit mesh: a vision-language model and a segmentation model are co-refined so their disagreements drive mutual correction, and a video diffusion model generates plausible articulation videos that serve only as motion hypotheses. Joint parameters are then fitted by minimizing 2D reprojection error, with the mesh geometry

What carries the argument

The load-bearing device is the explicit mesh M=(V,F) used as an intermediate representation. Surface connectivity gives spatially consistent part boundaries and reliable occlusion reasoning. Segmentation masks from multiple views are lifted onto the mesh through confidence-weighted evidence accumulation, then refined by graph-based label propagation that respects topology. For articulation, the mesh provides metric depth to unproject tracked 2D points into 3D rest-pose anchors, and later acts as a geometric referee: fitted revolute axes are penalized if they drive a child part into its parent, and the axis is softly blended with the principal direction of the adjacent part boundary. This mak

Load-bearing premise

The input mesh must be metrically scaled and geometrically faithful to the closed object, because joint fitting unprojects tracker points through mesh depth and validates axes against mesh geometry; any scale or shape error in the mesh propagates directly into joint position and type errors.

What would settle it

Reconstruct the same closed cabinet once with a metric multi-view mesh and once with a single-image generative mesh of unknown scale, then compare predicted joint axis positions to a measured ground truth; if the second version's axis position error jumps while the first stays low, the metric-scale mesh assumption is doing the load-bearing work, whereas similar accuracy would show robustness to scale-free inputs.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Digital twins of furniture can be built from rest-state product photos or scans, eliminating the need to observe or stage objects in an open state.
  • The number of movable parts is discovered automatically, so the pipeline applies to unseen objects without part-count priors.
  • Because the pipeline is mesh-based and reconstruction-agnostic, any source that produces a mesh — multi-view reconstruction, single-image generation, or an existing scan — can feed it without modification.
  • The output is usable in physics simulation and robotics pipelines: shell parts are solidified and drawer interiors are generated, yielding volumetric URDF-ready assets.
  • The same rest-state pipeline also handles partially open inputs with less reliance on the motion-hypothesis generator, since rest state is the harder setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the central claim holds, the practical bottleneck shifts from motion capture to metric-scale geometry: improving scale-accurate single-image reconstruction would directly strengthen rest-state joint estimation, while further video-generation fidelity may matter less once the mesh grounds the fit.
  • The co-refinement pattern — using disagreements between two pretrained models as a mutual correction signal rather than chaining them sequentially — is a transferable recipe for other ill-posed 3D tasks where multiple noisy priors must be reconciled against a shared geometric anchor.
  • A testable extension: feed the pipeline a mesh whose physical scale is deliberately wrong, e.g., uniformly scaled by 2x; if axis position error scales with the mesh, the metric-scale assumption is confirmed as load-bearing, and any future scale guarantee from single-image generators would unlock the method there.
  • The demonstrated categories are furniture-like objects with prismatic or revolute joints and fairly standard part vocabularies; extending to screw, ball, or multi-axis joints is an open direction the authors themselves flag, so the rest-state claim is best read as scoped to planar and linear articulation until then.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes Rest2Art, a framework for reconstructing articulated objects from a single closed-configuration mesh. The pipeline (i) co-refines part hierarchies from a VLM and per-view masks from SAM3 through an iterative mutual-correction loop, (ii) lifts validated masks onto the mesh via confidence-weighted evidence accumulation and graph-based label propagation, (iii) synthesizes articulation videos with a video diffusion model, extracts 2D tracks, and fits revolute/prismatic joint parameters using reprojection plus mesh-based geometric penalties, and (iv) volumetrizes parts for simulation. Evaluation is reported on ACD and MultiScan, with qualitative results on single-image and scanned-mesh inputs. The central claim is that rest-state articulation reconstruction with automatic part discovery is feasible and competitive with motion-observing baselines.

Significance. The core idea is timely and interesting: using an explicit mesh as a geometric anchor to cross-validate noisy foundation-model outputs, and using video-diffusion-generated motion only as a hypothesis source that is subsequently validated and refined by mesh geometry, is a sensible way to tackle an ill-posed problem. If the quantitative results hold under a fair and properly controlled protocol, the paper would be a meaningful step toward rest-state articulated-object reconstruction. The manuscript is also unusually detailed in its implementation parameters and ablation analyses (e.g., co-refinement convergence, video-generation robustness, mesh-backend sensitivity), which is a strength. However, the current evaluation has several issues that affect the strength of the central claims, particularly around metric scale, baseline input construction, and the real-world protocol.

major comments (4)
  1. [§3.3, §4.1, §4.5] The joint fitting procedure is metric-scale dependent: in §3.3 tracked pixels are unprojected using "the mesh provides metric depth at each tracked pixel," and the mesh-based penetration penalty operates in metric units. Consequently Axis Pos. (dm) and the geometric validation are meaningful only when the input mesh has correct scale. §4.1 states this only for the 2DGS backend, and Table 5 shows that segmentation and joint results degrade substantially with other backends (IoU 0.30 for 3DGS vs. 0.97 for 2DGS). Nevertheless, §4.5 and Fig. 9 claim generalization to single-image generation and scanned meshes without any scale calibration or quantitative metric-scale evaluation. Single-image generative models do not output metric-scale geometry, so the "input-agnostic" and "single closed configuration" claims, as stated, are supported only for metric multi-view reconstructions. Please either
  2. [§B.1, Table 2] The quantitative comparison with motion-observing baselines is confounded by the input-synthesis protocol. For ArtGS, REArtGS, REArtGS++, and Articulation in Motion, the two-state inputs are generated by the paper's own Wan2.2 pipeline, passed through Qwen-Image-Edit and MASt3R-SfM, and then manually filtered: §B.1 reports that "we manually verify the absence of camera drift and hallucination artifacts for every generated video on the ACD dataset, retaining 46 of 69 videos before running these multi-state baselines." This manual curation is not described for the proposed method's own video inputs, whose failures are instead handled automatically via truncation and fallback generation (§A.4). Moreover, these baselines are designed for real captures and are being evaluated on synthetic states generated by the proposed pipeline's components. Thus Table 2 does not establish the headline clai
  3. [§4.3, Table 3] The MultiScan real-world evaluation violates the rest-state assumption for the proposed method itself. The text states: "For single-state methods, including Articulate AnyMesh and ours, we use both available states as input and report the result with the lower reprojection error." This gives the method access to two configurations and permits cherry-picking the better output. Table 3 is the only real-world quantitative result, so the claim that the method works from rest-state observation in real scenes is not actually tested. Please report results using only the rest-state view, or if two states are allowed as a separate setting, report both states and the selection rule (e.g., average or worst-case) so the reader can disentangle the contribution of the second state.
  4. [Table 2] The main quantitative table contains an apparent data-integrity issue: in the ACD-HSSD block, Ours and Articulate AnyMesh both report Axis Ang. = 11.35 to two decimal places. Given the very different mechanisms of the two pipelines, this exact match is suspicious. Since Table 2 is the central evidence for the paper's main claim, please verify every entry and provide a corrected table if needed. In addition, no error bars or variance information are given anywhere in Tables 2–5, despite the stochastic nature of video generation and tracking; differences such as Ours vs. Singapo on the ACD-ABO block (4.78 vs. 0.13) may or may not be significant without per-scene standard errors.
minor comments (6)
  1. [Table 2] The row "Singapo + ⟳" is a modified baseline that uses the paper's co-refined hierarchy as input. This is an analysis of the co-refinement component, not an independent baseline, and should be clearly separated from the main comparison in the table or figure.
  2. [§4.3] Axis angular and positional errors are computed only for parts where the predicted joint type is correct. This conditional evaluation can inflate reported numbers; please also report the unconditional error or state why the conditional metric is preferred despite the selection bias.
  3. [§B.2] The MultiScan preprocessing retains "instances where the two states are not well-curated pairs" as real-world noise. This is a reasonable choice, but Table 3 reports only the evaluation coverage (Cov.) and not which objects or how many evaluable instances each method had. Reporting per-instance results or an intersection-of-coverage analysis would help the reader judge comparability.
  4. [§4.4 / Table 5] The mesh-backend ablation shows a large drop in segmentation IoU (0.97 for 2DGS to 0.30 for 3DGS) but only a modest change in joint direction error. A short explanation of why joint estimation is less sensitive than segmentation to the backend would be useful for interpreting the robustness claims.
  5. [§A.3] The hierarchy system prompt restricts part vocabulary to {base, door, knob, handle, drawer, tray}. The meta-prompt in §D.3 addresses broader categories, but the main paper does not clarify that the default system prompt is furniture-specific. Please state this limitation explicitly in the main text.
  6. [General] The paper would benefit from a short reproducibility statement and, ideally, release of code, given the large number of pretrained components and hand-tuned parameters. At minimum, the parameters in Table A should be cross-referenced from the main text.

Circularity Check

0 steps flagged

No significant circularity: the reconstruction and joint estimation are self-contained, with ground-truth joint parameters absent from all optimization steps.

full rationale

Rest2Art's derivation chain is not circular. Part segmentation (Sec. 3.2) fuses VLM hierarchy and SAM3 masks via cross-model co-refinement and mesh-grounded evidence accumulation; these are independent pretrained models, not outputs of the paper, and no ground-truth part labels are used. Articulation estimation (Sec. 3.3) uses a video diffusion model only to generate motion hypotheses, then fits rigid-body joint models to 2D tracks via reprojection (Eqs. 5-6), with mesh-based penetration and boundary-direction refinements as geometric constraints. Ground-truth joint parameters are never used in the optimization; the reported Axis Ang./Pos./Type are computed against held-out ACD and MultiScan ground truth. The paper's use of the same VDM to synthesize inputs for two-state baselines (Sec. B.1) is an evaluation-design choice that could affect comparative fairness, but it does not make the method's own predictions equivalent to its inputs. The metric-scale dependence on the 2DGS backend (Sec. 4.1) is a stated assumption and a correctness/brittleness risk for single-image inputs, not a circularity: the network weights of Wan2.2, CoTracker3, GPT-5.2, and SAM3 are external, and no self-citation is load-bearing. The paper also candidly notes rest-state interior inference is ill-posed and limits its mesh completion to volumetrization (Sec. A.6), which is a limitation, not circularity. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is relabeled as unification.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 0 invented entities

The paper is an empirical pipeline; no new physical or mathematical entities are introduced. The central claim rests on the reliability of several pretrained models (VLM, SAM, VDM, tracker) and on the metric fidelity of the input mesh. Hyperparameters are hand-set engineering choices rather than fitted scientific parameters.

free parameters (7)
  • Top-K views per part = not stated
    Sec. 3.2 selects top-K views but Table A never lists K; controls evidence accumulation in Eq. (2).
  • Visibility truncation threshold tau_vis = 0.5
    Table A / A.5; determines how many frames of synthesized video are used for joint fitting.
  • Number of video seeds R = 2
    Table A; joint estimates are refit from two seeds; fallback adds two more videos if static.
  • Max co-refinement rounds Rmax = 2
    Table A; 84.1% of scenes converge; raising to 20 improves to 92.8% (Sec. C.1), so the cap trades accuracy for cost.
  • Mask overlay alpha = 0.35
    A.4; blending of part masks into the video-diffusion conditioning image; ablation shows overlay affects hallucination rate.
  • Solidification thickness delta = 0.015 m
    A.6; inward offset thickness for volumetric mesh completion; adaptively clamped per-vertex.
  • Camera radius for standalone mesh inputs = 4.5
    Table A; fixed camera placement for mesh-only inputs; affects viewpoint coverage and view selection.
axioms (7)
  • domain assumption Input mesh reconstructed by 2DGS preserves metric scale consistent with ground truth.
    Stated in Sec. 4.1; joint-position errors and track unprojection in Sec. 3.3 use metric depth from the mesh.
  • domain assumption GPT-5.2 can propose a correct part hierarchy and instance counts from closed-state views.
    Used in Sec. 3.2 co-refinement; Table 4 shows VLM-only tree accuracy is only 62.3%, so the hierarchy is often wrong before correction.
  • domain assumption SAM3 can segment prompted parts in individual views.
    Used in Sec. 3.2; SAM3-only part count accuracy is 82.6% (Table 4), so masks are noisy and need cross-view validation.
  • domain assumption Wan2.2+VBVR video diffusion can imagine articulation whose motion type and direction are approximately correct.
    Central to Sec. 3.3; the paper requires only approximate type/direction, but hallucinations still occur (2/69 corrupted videos in Sec. C.4).
  • domain assumption CoTracker3 tracks feature points sufficiently accurately across synthesized video frames.
    Used for 2D tracks in Sec. 3.3; no independent validation of tracker accuracy on generated videos.
  • domain assumption Joints are limited to prismatic and revolute types.
    Stated in Sec. 3.1; this restricts applicability to furniture-like mechanisms and excludes more general articulations.
  • domain assumption Ground-truth annotations in ACD and MultiScan are correct.
    Evaluation relies on them; no manual re-annotation described.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Articulated Object Reconstruction from Rest-State Observation." pith.science (2026). https://pith.science/paper/N2CGE635

@misc{pith2026260727749,
  author       = {Pith},
  title        = {Pith review of: Articulated Object Reconstruction from Rest-State Observation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2CGE635}},
  note         = {Machine review of arXiv:2607.27749}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Building interactive digital twins requires recovering both 3D geometry and the kinematic structures that govern how objects articulate. Yet existing methods for articulated object reconstruction require explicitly observable motion from multiple articulation states. We introduce a rest-state formulation that reconstructs articulated objects from a single closed configuration, an inherently ill-posed setting where geometry, semantics, and motion priors compensate for the absence of motion cues. Our framework adopts an explicit mesh as an intermediate representation for cross-model verification and fusion, reconciling noisy outputs from vision-language and segmentation models into spatially consistent part structures. To estimate joint parameters without observed motion, we use a video diffusion model to synthesize articulation hypotheses and validate them through geometric consistency. Our approach achieves accurate part decomposition and physically plausible articulation, performing competitively with motion-observing reconstruction-based, generation-based, and modular pretrained-model baselines.

Figures

Figures reproduced from arXiv: 2607.27749 by Daeun Lee, Haebeom Jung, Jaeah Lee, Jaesik Park, Woosung Kim.

Figure 1
Figure 1. Figure 1: We propose Rest2Art, an approach for reconstructing articulated objects with part-level geometry and joint parameters from rest-state observations alone. Abstract. Building interactive digital twins requires recovering both 3D geometry and the kinematic structures that govern how objects articu￾late. Yet existing methods for articulated object reconstruction require explicitly observable motion from multip… view at source ↗
Figure 2
Figure 2. Figure 2: Naïvely applying pretrained models to rest-state articulated objects. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Method overview. Rest2Art reconstructs articulated objects from rest-state inputs by co-refining part hierarchies and masks, lifting evidence onto the mesh, syn￾thesizing videos for joint estimation, and completing interior geometry. 3.1 Problem Statement We focus on multi-part openable objects, such as furniture, whose closed con￾figuration can be clearly defined. Formally, given a rest-state object mesh,… view at source ↗
Figure 4
Figure 4. Figure 4: Co-refinement of hierarchy and part masks. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of part segmentation on ACD dataset [15]. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of joint estimation on ACD dataset [15]. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative ablation on mesh reconstruction methods. Original Masked Input image Wan2.2 Wan2.2 + VBVR LoRA [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative ablation on video generation. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Generalization to Diverse Inputs. Our method applies to single-image inputs and scanned meshes, in addition to multi-view captures. 4.5 Generalization to Diverse Inputs Since our framework is input-agnostic, it extends beyond multi-view captures. For single-image inputs, we replace the multi-view reconstruction stage with an image-to-3D generative model [45, 54]; for scanned meshes, the pipeline applies di… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 7 linked inside Pith

  1. [1]

    Ai, H., Chang, W., Jiao, J., Leonardis, A., Ofek, E.: Articulation in motion: Prior- free part mobility analysis for articulated objects by dynamic-static disentangle- ment. In: Int. Conf. Learn. Represent. (2026)

    2026

  2. [2]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, ...

    Pith/arXiv arXiv 2025

  3. [3]

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala, K.V., Khedr, H., Huang, A., Lei, J., Ma, T., Guo, B., Kalla, A., Marks, M., Greer, J., Wang, M., Sun, P., Rädle, R., Afouras, T., Mavroudi, E., Xu, K., Wu, T.H., Zhou, Y., Momeni, L., Hazra, R., Ding, S., Vaze, S., Porcher, F., Li, F., Li, S., Kamath, A., Cheng, H.K., ...

    2026

  4. [4]

    arXiv preprint arXiv:1512.03012 (2015)

    Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015)

    Pith/arXiv arXiv 2015

  5. [5]

    In: Proceedings of Robotics: Science and Systems (2024)

    Chen, Z., Walsman, A., Memmel, M., Mo, K., Fang, A., Vemuri, K., Wu, A., Fox, D., Gupta, A.: Urdformer: A pipeline for constructing articulated simulation environments from real-world images. In: Proceedings of Robotics: Science and Systems (2024)

    2024

  6. [6]

    In: IEEE Conf

    Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: ABO: Dataset and benchmarks for real-world 3D object understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21126–21136 (2022)

    2022

  7. [7]

    Contributors, L.: Lightx2v: Light video generation inference framework.https: //github.com/ModelTC/lightx2v(2025)

    2025

  8. [8]

    In: IEEE Conf

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5828–5839 (2017)

    2017

  9. [9]

    In: IEEE Conf

    Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of an- notated 3D objects. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13142–13153 (2023)

    2023

  10. [10]

    Dharmarajan, K., Huang, W., Wu, J., Fei-Fei, L., Zhang, R.: Dream2flow: Bridging video generation and open-world manipulation with 3d object flow. In: Int. Conf. on Robotics and Automation (ICRA) (2026)

    2026

  11. [11]

    In: International Conference on 3D Vision (3DV)

    Duisterhof, B.P., Zust, L., Weinzaepfel, P., Leroy, V., Cabon, Y., Revaud, J.: Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. In: International Conference on 3D Vision (3DV). pp. 1–10. IEEE (2025)

    2025

  12. [12]

    In: IEEE Conf

    Guo, J., Xin, Y., Liu, G., Xu, K., Liu, L., Hu, R.: Articulatedgs: Self-supervised digital twin modeling of articulated objects using 3d gaussian splatting. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 27144–27153 (2025) Articulated Object Reconstruction from Rest-State Observation 17

    2025

  13. [13]

    In: IEEE Conf

    Heppert, N., Irshad, M.Z., Zakharov, S., Liu, K., Ambrus, R.A., Bohg, J., Valada, A., Kollar, T.: Carto: Category and joint agnostic reconstruction of articulated objects. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21201–21210 (2023)

    2023

  14. [14]

    In: SIGGRAPH Conference Papers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geomet- rically accurate radiance fields. In: SIGGRAPH Conference Papers. Association for Computing Machinery (2024).https://doi.org/10.1145/3641519.3657428

    arXiv 2024

  15. [15]

    In: IEEE/CVF Winter Conference on Applications of Computer Vision

    Iliash, D., Jiang, H., Zhang, Y., Savva, M., Chang, A.X.: S2o: Static to open- able enhancement for articulated 3d objects. In: IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6785–6795 (2026)

    2026

  16. [16]

    In: IEEE Conf

    Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5616–5626 (2022)

    2022

  17. [17]

    Karaev, N., Makarov, Y., Wang, J., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker3: Simpler and better point tracking by pseudo-labelling real videos. In: Int. Conf. Comput. Vis. pp. 6013–6022 (2025)

    2025

  18. [18]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  19. [19]

    In: IEEE Conf

    Khanna*, M., Mao*, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16384–16393 (June 2024)

    2024

  20. [20]

    Kijai: Vbvr lora for wan2.2 i2v.https://huggingface.co/Kijai/WanVideo_comfy (2025), accessed: 2026-03-08

    2025

  21. [21]

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015–4026 (2023)

    2023

  22. [22]

    Le, L., Xie, J., Liang, W., Wang, H.J., Yang, Y., Ma, Y.J., Vedder, K., Krishna, A., Jayaraman, D., Eaton, E.: Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model. In: Int. Conf. Learn. Represent. vol. 2025, pp. 17578–17602 (2025)

    2025

  23. [23]

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: Eur. Conf. Comput. Vis. pp. 71–91. Springer (2024)

    2024

  24. [24]

    Li, R., Zheng, C., Rupprecht, C., Vedaldi, A.: Dragapart: Learning a part-level motion prior for articulated objects. In: Eur. Conf. Comput. Vis. pp. 165–183. Springer (2024)

    2024

  25. [25]

    In: IEEE Conf

    Li, X., Wang, H., Yi, L., Guibas, L.J., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3706–3715 (2020)

    2020

  26. [26]

    Liu, J., Iliash, D., Chang, A., Savva, M., Mahdavi Amiri, A.: Singapo: Single image controlled generation of articulated parts in objects. In: Int. Conf. Learn. Represent. vol. 2025, pp. 97511–97532 (2025)

    2025

  27. [27]

    Liu, J., Mahdavi-Amiri, A., Savva, M.: PARIS: Part-level reconstruction and mo- tion analysis for articulated objects. In: Int. Conf. Comput. Vis. pp. 352–363 (2023)

    2023

  28. [28]

    Liu, M., Uy, M.A., Xiang, D., Su, H., Fidler, S., Sharp, N., Gao, J.: Partfield: Learning3dfeaturefieldsforpartsegmentationandbeyond.In:Int.Conf.Comput. Vis. pp. 9704–9715 (2025)

    2025

  29. [29]

    In: IEEE Conf

    Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21736–21746 (2023) 18 D. Lee et al

    2023

  30. [30]

    Liu,Y.,Jia,B.,Lu,R.,Ni,J.,Zhu,S.C.,Huang,S.:Buildinginteractablereplicasof complex articulated objects via gaussian splatting. In: Int. Conf. Learn. Represent. (2025)

    2025

  31. [31]

    Ma, Z., Yue, Y., Gkioxari, G.: Find any part in 3d. In: Int. Conf. Comput. Vis. pp. 7818–7827 (2025)

    2025

  32. [32]

    Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: MultiScan: Scalable RGBD scanning for 3D environments with articulated objects. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 9058–9071 (2022)

    2022

  33. [33]

    Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3d point clouds. In: Int. Conf. on 3D Vision (3DV). pp. 992–1002. IEEE (2021)

    2021

  34. [34]

    In: IEEE Conf

    Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object un- derstanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 909–918 (2019)

    2019

  35. [35]

    NVIDIA: Isaac Sim,https://github.com/isaac-sim/IsaacSim

  36. [36]

    Odin, L.: Qwen-image-edit-2511 multiple angles lora.https://huggingface.co/ fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA(2025), accessed: 2026-03-08

    2025

  37. [37]

    OpenAI: GPT-5 system card.https://cdn.openai.com/gpt-5-system-card.pdf (2025), accessed: 2026-05-20

    2025

  38. [38]

    In: Conference on Robot Learning (CoRL) (2025)

    Qiu, X., Yang, J., Wang, Y., Chen, Z., Wang, Y., Wang, T.H., Xian, Z., Gan, C.: Articulate AnyMesh: Open-vocabulary 3D articulated objects modeling. In: Conference on Robot Learning (CoRL) (2025)

    2025

  39. [39]

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. In: Int. Conf. Learn. Represent. vol. 2025, pp. 28085–28128 (2025)

    2025

  40. [40]

    arXiv preprint arXiv:2007.08501 (2020)

    Ravi,N.,Reizenstein,J.,Novotny,D.,Gordon,T.,Lo,W.Y.,Johnson,J.,Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)

    Pith/arXiv arXiv 2007

  41. [41]

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)

    2024

  42. [42]

    arXiv preprint arXiv:1906.05797 (2019)

    Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur- Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.:...

    Pith/arXiv arXiv 1906

  43. [43]

    In: IEEE Conf

    Sun, C., Choe, J., Loop, C., Ma, W.C., Wang, Y.C.F.: Sparse voxels rasteriza- tion: Real-time high-fidelity radiance field rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16187–16196 (2025)

    2025

  44. [44]

    arXiv preprint arXiv:2408.13679 (2024)

    Tang, G., Zhao, W., Ford, L., Benhaim, D., Zhang, P.: Segment any mesh. arXiv preprint arXiv:2408.13679 (2024)

    Pith/arXiv arXiv 2024

  45. [45]

    In: IEEE Conf

    Team, S.D., Chen, X., Chu, F.J., Gleize, P., Liang, K.J., Sax, A., Tang, H., Wang, W., Guo, M., Hardin, T., Li, X., Lin, A., Liu, J., Ma, Z., Sagar, A., Song, B., Wang, X., Yang, J., Zhang, B., Dollár, P., Gkioxari, G., Feiszli, M., Malik, J.: Sam 3d: 3dfy anything in images. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7220–7232 (2026) Articulated Obje...

    2026

  46. [46]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, T., Wang, T., Gui, T., Weng, T., Shen, T., Lin, W., Wang, W., Wang, W., Zhou, W.,...

    Pith/arXiv arXiv 2025

  47. [47]

    arXiv preprint arXiv:2602.20159 (2026)

    Wang, M., Wang, R., Lin, J., Ji, R., Wiedemer, T., Gao, Q., Luo, D., Qian, Y., Huang, L., Hong, Z., et al.: A very big video reasoning suite. arXiv preprint arXiv:2602.20159 (2026)

    arXiv 2026

  48. [48]

    In: IEEE Conf

    Weng, Y., Wen, B., Tremblay, J., Blukis, V., Fox, D., Guibas, L., Birchfield, S.: Neural implicit representation for building digital twins of unknown articulated objects. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3141–3150 (2024)

    2024

  49. [49]

    In: Conference on Robot Learning (CoRL)

    Werby, A., Büchner, M., Röfer, A., Huang, C., Burgard, W., Valada, A.: Artic- ulated object estimation in the wild. In: Conference on Robot Learning (CoRL). vol. 2 (2025)

    2025

  50. [50]

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., L...

    Pith/arXiv arXiv 2025

  51. [51]

    In: IEEE Conf

    Wu, D., Liu, L., Huang, A., Liu, Y., Yu, Q., Liu, S., Song, L., Lu, C.: ReArtGS++: Generalizable articulation reconstruction with temporal geometry constraint via planar gaussian splatting. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1177– 1186 (2026)

    2026

  52. [52]

    Wu, D., Liu, L., Linli, Z., Huang, A., Song, L., Yu, Q., Wu, Q., Lu, C.: ReArtGS: Reconstructing and generating articulated objects via 3D gaussian splatting with geometric and motion constraints. In: Adv. Neural Inform. Process. Syst. vol. 38, pp. 102889–102915 (2026)

    2026

  53. [53]

    In: IEEE Conf

    Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11097–11107 (June 2020)

    2020

  54. [54]

    In: IEEE Conf

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21469–21480 (2025)

    2025

  55. [55]

    In: Brit

    Xue, H., Liu, L., Xu, W., Fu, H., Lu, C.: Omad: Object model with articulated deformations for pose estimation and retrieval. In: Brit. Mach. Vis. Conf. (2021)

    2021

  56. [56]

    Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Int. Conf. Comput. Vis. pp. 12–22 (2023)

    2023

  57. [57]

    In: SIGGRAPH Asia Conference Papers

    Yuan, S., Shi, R., Wei, X., Zhang, X., Su, H., Liu, M.: Larm: A large articulated object reconstruction model. In: SIGGRAPH Asia Conference Papers. pp. 1–12 (2025)

    2025

  58. [58]

    In: IEEE Conf

    Zhang, C., Lee, G.H.: Iaao: Interactive affordance learning for articulated objects in 3d environments. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12132–12142 (2025) Articulated Object Reconstruction from Rest-State Observation 1

    2025

  59. [59]

    Zhao, M., Weng, Y., Bauer, D., Song, S.: Real2code: Reconstruct articulated ob- jects via code generation. In: Int. Conf. Learn. Represent. vol. 2025, pp. 668–686 (2025)

    2025

  60. [60]

    ‘json

    Zhou, Y., Gu, J., Li, X., Liu, M., Fang, Y., Su, H.: Partslip++: Enhancing low- shot 3d part segmentation via multi-view instance segmentation and maximum likelihood estimation. Int. Conf. Comput. Vis. Worksh. (2025) A Implementation Details ThissectionprovidesadditionalimplementationdetailsforeachstageofRest2Art. ◦Section A.1 lists the key implementation...

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.