Pith. sign in

REVIEW 2 major objections 6 minor 41 references

Oriented bounding boxes plus two structure losses let Gaussian splatting recover clean articulated parts and motion from images alone.

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 ↗

T0 review

2026-07-30 18:00 UTC pith:RKHD4JHQ

load-bearing objection Solid articulated-3DGS recipe: OBB fitting + SAT contact actually fix the shape–motion mess photometric methods leave, especially in sparse views. the 2 major comments →

arxiv 2607.26889 v1 pith:RKHD4JHQ submitted 2026-07-29 cs.GR cs.CVcs.RO

StructureGS: Structure-aware Gaussian Splatting for Articulated Object Reconstruction

Gahye Lee , Gyoonseo Kim , Wonjong Jang , Jooeun Son , Seungyong Lee This is my paper
classification cs.GR cs.CVcs.RO
keywords articulated object reconstruction3D Gaussian Splattingoriented bounding boxesstructure-aware losspart decompositionspatial coherencestructural connectivitysparse-view reconstruction
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.

Reconstructing an object that can open, slide, or fold is hard because the same pixels must explain geometry, appearance, and how parts move. Pure image-matching optimizers often smear geometry across part boundaries or invent wrong joints. StructureGS represents each rigid part by canonical 3D Gaussians inside an oriented bounding box, then adds two explicit losses: one that forces each box to tightly cover its part without excess empty space, and one that keeps adjacent parts in contact via a differentiable box-distance test. Together with ordinary rendering loss, these constraints disentangle shape from motion, yielding sharp part boundaries and accurate revolute or prismatic joints even from only ten views or noisy real phone captures. The practical payoff is digital twins that can be animated and interacted with, not just pretty renderings of the whole object.

Core claim

When part-wise 3D Gaussian Splatting is optimized under OBB-based part-fitting and part-contact losses (plus a light extent regularizer), the joint estimation of per-part geometry, appearance, and rigid motion produces well-separated parts and accurate articulation parameters that photometric supervision alone cannot reliably recover, including under sparse multi-view input.

What carries the argument

Structure-aware losses on oriented bounding boxes: part fitting (coverage plus margin on rendered structure images) for spatial coherence, and part contact (squared hinge on Separating-Axis-Theorem distances between base and movable OBBs) for structural connectivity; both act on the same affine/rigid transforms that place the canonical Gaussians.

Load-bearing premise

That simple oriented boxes plus contact only between a static base and each movable part are enough structural prior to keep the optimizer from collapsing or merging parts when silhouette change is weak or true contact topology is more complex.

What would settle it

On scenes with weak visual motion cues (e.g., the oven door case) or multi-hop nested contacts, measure whether part-level Chamfer distance and joint-axis error stay low after optimization; systematic part collapse or large residual separation would falsify the claim that the OBB losses suffice.

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

If this is right

  • Sparse-view and real-phone capture pipelines can produce usable articulated digital twins without dense multi-view rigs.
  • Part-wise animation and interaction become reliable once geometry and joint parameters are no longer entangled by pure photometric loss.
  • The same OBB-plus-contact regularizer can be dropped into other Gaussian or neural articulated reconstructors that currently rely only on image matching.
  • Multi-part objects (4–7 moving parts) remain tractable under the same base–movable contact graph used for simpler two-part cases.

Where Pith is reading between the lines

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

  • When silhouette change is weak, adding semantic features (as the authors briefly suggest) may be necessary before the geometric prior can lock parts; pure structure losses are not a complete substitute for appearance signal.
  • The method’s tolerance of coarse VGGT/DBSCAN initialization implies that structure losses can act as a corrector for imperfect motion segmentation upstream of reconstruction.
  • Extending the contact graph beyond star topology (base to each movable) would be a natural next test for objects whose parts touch each other rather than only the base.

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

2 major / 6 minor

Summary. StructureGS reconstructs articulated objects from multi-view RGB of two articulation states by representing each part as canonical 3D Gaussians plus an oriented bounding box (affine T0_k) and a rigid articulation transform T1_k. Beyond photometric loss, it adds OBB-driven structure-aware terms: a part-fitting loss (coverage + margin on rendered OBB structure images) for spatial coherence, a SAT-based part-contact loss that penalizes separation of base–movable pairs, and an extent regularizer. The paper reports SOTA geometry, motion, and novel-view metrics on PARIS (full and 10-view sparse), DTA multi-part, ArtGS-Multi (supp.), and sparse real smartphone captures, with ablations and an honest weak-silhouette failure case.

Significance. Shape–motion entanglement under photometric-only supervision is a genuine bottleneck for part-wise articulated reconstruction, especially under sparse views. The OBB reparameterization plus differentiable fitting/contact losses is a clear, implementable idea that yields large, consistent gains on standard benchmarks (Tables 2–3; Figs. 5–9), including practical 10-view phone data. Strengths include multi-benchmark evaluation, loss and sparsity ablations, robustness checks on noisy OBB init (supp.), and explicit disclosure of a failure mode (Fig. 10). If the claims hold under the stated two-state, known-K regime, the work is a solid incremental advance for interactive digital twins and embodied AI perception.

major comments (2)
  1. [Sec. 4.1, Eq. (4); Table 3; supp. ArtGS-Multi] Sec. 4.1, Eq. (4): the contact graph E is defined only as pairs (P0, Pk) between the static base and each movable part. The multi-part SOTA claim (DTA; ArtGS-Multi in supp., 4–7 parts) therefore rests on a star topology. Nested or part–part contacts (e.g., stacked drawers, linked mechanisms) are not modeled; SAT-on-OBB may still allow overlap but never enforces contact between movable parts. Please either (i) extend E to a user-/init-supplied adjacency (or learn it) and re-evaluate multi-part scenes, or (ii) clearly bound the multi-part claim to base-attached kinematics and report failure modes when that assumption breaks.
  2. [Sec. 3; Sec. 4.2; Sec. 6; supp. §8.3] Sec. 3–4 and supp. init: K (number of parts) and the base/movable partition are treated as known inputs (DBSCAN cluster count / expected part count; E built from P0). This is common but load-bearing for the “automatic” part decomposition narrative in the abstract and intro. State this assumption up front in the problem setup and limitations; if K is misspecified, quantify degradation (merge/split) rather than only showing robustness to noisy box geometry with correct K.
minor comments (6)
  1. [Table 1; Sec. 2] Table 1 “Structure-awareness” is binary and self-favoring; a short note on what counts as structural (vs. mobility/clustering priors in ScrewSplat, ArtGS, SPLART) would avoid appearing to dismiss related work.
  2. [Sec. 4.2; supp. §8.1] Loss schedule and weights (λ_image=10, λ_coverage=100, λ_margin=50, contact/ext off at 1500) are given only in the supplement. A one-sentence pointer in Sec. 4.2 and a brief sensitivity note would help reproducibility in the main text.
  3. [Fig. 8; Sec. 5.4] Fig. 8 ablation: the large CD-m drop after L_ext is attributed mainly to Blade; report per-scene or median metrics so one scene does not dominate the narrative that extent regularization is generally essential.
  4. [Sec. 5.1; supp. §8.4] Joint-type rule (supp.: revolute if rotation angle ≥5°) should be mentioned briefly in the main kinematics evaluation paragraph so axis/geodesic metrics are interpretable without the supplement.
  5. [Sec. 3; Fig. 2] Minor typos/notation: “att=0” spacing; “w/ L_contact” figure callout; consistent use of T^k_0 vs T_0^k; arXiv IDs in related work are fine but check ScrewSplat/SPLART citation completeness for camera-ready.
  6. [Sec. 5.1; Sec. 5.3] Real-world pipeline depends on COLMAP+SAM masks; a short note on mask error sensitivity (or qualitative failure when masks leak) would strengthen Sec. 5.3.

Circularity Check

0 steps flagged

No significant circularity: empirical reconstruction method with independent losses and external benchmarks.

full rationale

StructureGS is an empirical graphics method, not a first-principles derivation. The load-bearing claim is that OBB-based part-fitting and part-contact losses, added to photometric 3DGS optimization, improve part decomposition and motion estimation versus photometric-only baselines. Those losses are regularizers (Eqs. 2–6); evaluation uses independent metrics—Chamfer distances on sampled geometry, joint angular/position/geodesic errors, and novel-view PSNR/SSIM/LPIPS—on held-out views and external benchmarks (PARIS, DTA, ArtGS-Multi, real captures). Ablations (Fig. 8) show removing structure terms worsens CD and motion, so reported gains are not tautological restatements of the objective. Initialization (VGGT + DBSCAN OBBs) is coarse and refined jointly; it does not force the final metrics by construction. Citations (3DGS, SAT/Ericson, ScrewSplat, ArticulatedGS, PARIS) are standard prior art or baselines, not self-citation uniqueness theorems that close the argument. No step reduces a claimed prediction to a fitted input or a definitional identity. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

Load-bearing content is an engineering prior stack: rigid parts, two discrete articulation states, OBB as structure proxy, base-centric contact graph, and many loss/schedule knobs. No new physical entities; contribution is the loss-guided GS pipeline and empirical performance under those assumptions.

free parameters (6)
  • Loss weights λ_image, λ_coverage, λ_margin, λ_contact, λ_ext = 10 / 100 / 50 / 0.1 / 0.01
    Hand-chosen defaults in supp (10, 100, 50, 0.1, 0.01); central quality depends on this balance between photometry and structure.
  • Optimization schedule (L_image on at 400; L_contact & L_ext off at 1500; 10k iters) = on@400; structure terms off@1500; 10000 iters
    Progressive enabling/disabling of terms is a free procedural choice that shapes convergence and final part splits.
  • Gaussians per part initialization count = 10000
    Fixed at 10,000 random canonical samples per part; affects capacity and densification trajectory.
  • Revolute vs prismatic angle threshold = 5 degrees
    Joint type from rotation angle of T1_k with 5° cutoff; changes reported kinematics metrics.
  • Motion-cue threshold τ_motion and DBSCAN clustering settings for OBB init
    Separate static/moving points and part hypotheses before fitting initial OBBs; paper stresses robustness but values are free init hyperparameters.
  • Number of parts K and base-movable adjacency set E = scene-dependent K; E={(0,k)}
    Model assumes known/chosen part count and contacts only between base and each movable part; not inferred end-to-end from images alone.
axioms (5)
  • domain assumption Object is a composition of K rigid parts with one fixed base; motion between two observed states is a single rigid transform per movable part.
    Stated in Sec. 3 opening; excludes non-rigid, multi-state continuous trajectories, and unknown part counts.
  • ad hoc to paper An oriented bounding box induced by affine map T0_k is a sufficient proxy for part spatial extent to enforce coherence and contact.
    Core design in Secs. 3.1–3.2; Gaussians recover fine shape, but structure losses act on OBBs.
  • domain assumption Physically plausible articulation implies non-separation of base and each movable part under SAT OBB distance (overlap allowed, separation penalized).
    Eq. 4 and paragraph on nested drawers; contact graph is simplified to base links.
  • domain assumption Multi-view RGB of exactly two configurations plus off-the-shelf poses/masks (and VGGT/Pi3 points for init) suffice to recover geometry and kinematics.
    Problem setup Sec. 3 and real-world protocol Sec. 5.1; COLMAP/SAM noise acknowledged.
  • standard math Separating Axis Theorem distance between convex OBBs is a valid differentiable contact metric for training.
    Sec. 3.2 cites classic collision geometry; used as d(Pi,Pj)=max_a da.
invented entities (3)
  • StructureGS part model (canonical tanh-bounded Gaussians + OBB affine T0 + rigid T1) no independent evidence
    purpose: Disentangle intrinsic part geometry from pose/extent and enable joint photometric+structure optimization.
    Representational packaging specific to this paper; builds on 3DGS but ties membership and motion to OBB parameters.
  • Part fitting loss via rendered OBB structure images (L_coverage + L_margin) no independent evidence
    purpose: Enforce spatial coherence by covering the object mask tightly without excess exterior volume.
    Novel supervisory signal constructed in Sec. 4.1; validated only inside this paper’s ablations.
  • Part contact loss using SAT OBB distances no independent evidence
    purpose: Penalize separation between adjacent parts during optimization.
    Applies known SAT math as a new training loss for articulated GS; no external physical validation beyond recon metrics.

pith-pipeline@v1.2.0-daily-grok45 · 24125 in / 3835 out tokens · 74531 ms · 2026-07-30T18:00:31.320296+00:00 · methodology

0 comments
read the original abstract

Reconstructing articulated objects with multiple movable parts is essential for understanding object structure and enabling physical interaction. However, this reconstruction task poses significant challenges due to the entanglement of geometry, appearance, and motion parameters during optimization. Existing methods rely primarily on photometric supervision, which commonly fails to disentangle these interdependent components, resulting in poor part decomposition with blurred boundaries and geometric artifacts. To address this limitation, we introduce StructureGS, a reconstruction framework for articulated objects that integrates structure-aware guidance into 3D Gaussian Splatting. Our approach leverages oriented bounding boxes of object parts to enforce two key structural properties: spatial coherence, which constrains each part's geometry to remain compact and spatially coherent within its designated region, and structural connectivity, which enforces physically plausible contact relationships between adjacent parts. These properties are realized through structure-aware losses that inject explicit structural constraints into the optimization process. Extensive experiments demonstrate that our method achieves state-of-the-art performance in articulated object reconstruction, producing high-quality results with well-defined part geometries.

Figures

Figures reproduced from arXiv: 2607.26889 by Gahye Lee, Gyoonseo Kim, Jooeun Son, Seungyong Lee, Wonjong Jang.

Figure 1
Figure 1. Figure 1: StructureGS. (a) Each part is represented by an explicit OBB (structure) and 3D Gaussians (geometry), with prismatic or revolute motion parameters. (b) Structure￾aware optimization yields cleaner part decomposition and more accurate motion esti￾mation than a photometric-only method (ScrewSplat [12]). on understanding how parts are spatially structured and move relative to each other. This entanglement ofte… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of StructureGS. (a) Each part Pk is represented by canonical Gaus￾sians Gk, which are transformed to two states: an initial state (t = 0) and an articulated state (t = 1). (b) For each state, the transformed Gaussians are rendered with the 3DGS renderer, and photometric and structure-aware losses are computed to jointly optimize Gk, T k 0 , and T k 1 . 3 Structure-Aware Gaussian Splatting Given mu… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of part fitting loss. (a) P1’s OBB is fully inside the part (margin loss = 0), but does not cover the entire part, incurring coverage loss. (b) P1’s OBB fully covers the part (coverage loss = 0), but includes regions outside the part, incurring margin loss. (c) Optimizing both losses yields a tight OBB. the structure images, we formulate the part fitting loss with two complementary objectives:… view at source ↗
Figure 4
Figure 4. Figure 4: Structure-aware optimization over iterations. Part-wise 3D Gaussian splats constrained by OBBs progressively align with their corresponding object parts. Extent regularization To encourage the OBB scale components to adapt to the underlying part geometry, we introduce an axis-weighted scale regularization: \mathcal {L}_{\text {ext}} = -\lambda _{\text {ext}} \sum _{k=0}^{K-1} \sum _{i=1}^{3} w_k^{(i)} \log… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on PARIS and DTA datasets under full- and sparse￾view settings. Reconstructed geometry, fitted OBBs with per-part 3DGS, and estimated joint axes (red) are shown; gray denotes static parts, while blue and green indicate movable parts. Additional results are provided in the supplementary material. Initial state Articulated state ScrewSplat Ours ArticulatedGS Ours Initial state Articula… view at source ↗
Figure 6
Figure 6. Figure 6: Articulation animation on PARIS real data from initial to articulated states. fair comparison, we adapt ScrewSplat to the two-state setting according to its original protocol, evaluating multiple parsimony weights per scene to report the best achievable result. Evaluation metrics We comprehensively evaluate our method across three as￾pects: geometry, kinematics, and rendering. Following [7,18], geometric a… view at source ↗
Figure 7
Figure 7. Figure 7: Reconstruction of real-world articulated objects using 10 views. Blue/green Gaussian splats denote static/dynamic parts, respectively. Joint axes and estimated OBBs of our method are shown as red arrows and black wireframes, respectively. 5.3 Real-world evaluation We further evaluate our method’s generalization under sparse and noisy real￾world conditions using only 10 views per state. As shown in [PITH_F… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation on structural losses. Left: Visual comparison of the effect of each loss term on reconstruction quality. Right: Quantitative results for each loss configuration [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Robustness to sparse observations. Angular error and Chamfer distance on the PARIS benchmark under varying numbers of training views. Our method remains stable as views become sparse, while ArticulatedGS [7] degrades sharply. adding Lext mainly comes from the Blade scene. There, the regularizer helps the OBB recover the full movable-part extent, rather than fitting only to the exposed tip where silhouette … view at source ↗
Figure 10
Figure 10. Figure 10: Failure case. (a) Target object in the initial and articulated states. (b) Re￾constructed 3DGS in the corresponding states, with parts color-coded. (c) Zoomed-in views showing part collapse in the articulated state. 6 Conclusion In this work, we presented StructureGS, a framework for articulated object re￾construction that achieves clean part decomposition through structure-aware guidance. By representing… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of the error maps and intermediate terms of the part fitting loss. M denotes the binary ground-truth object mask, \protect \mathbf {I}_{\text {struct}}^{k} the occupancy map of the k-th part OBB, and \protect \mathbf {I}_{\text {struct}}^{\text {all}} their union. (a) and (b) illustrate two example cases where the coverage loss and the margin loss are non-zero, respectively. In (a), the erro… view at source ↗
Figure 12
Figure 12. Figure 12: Left: Default initial OBBs obtained from motion-based point clustering, where point colors indicate clustered part points. Middle: OBB optimization results over iterations under clustering noise, point cloud noise, and an alternative point cloud estimator. Right: Fitting error curves over optimization iterations for the default and perturbed settings. estimator (VGGT) with a recent alternative method (Pi3… view at source ↗
Figure 13
Figure 13. Figure 13: Further qualitative comparison on the PARIS dataset [18] under full-view and sparse-view settings. The left column shows the articulated objects in two states. The remaining columns visualize reconstruction and articulation estimation results of different methods (PARIS [18], ScrewSplat [12], ArticulatedGS [7], and ours). Dynamic parts are shown in blue and static parts in gray, while red arrows indicate … view at source ↗
Figure 14
Figure 14. Figure 14: Gallery of part-wise reconstructed meshes across different object categories [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Animation results on real-world objects. For each example, the left two columns show the input image examples at the initial and articulated states, respec￾tively. The remaining columns show rendered frames from the resulting animation, where the reconstructed part-level Gaussians are transformed according to the esti￾mated articulation parameters [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: OBB and 3DGS fitting results on the PARIS dataset. Columns correspond to different scenes from the PARIS benchmark. Rows show the full reconstruction (“All”) and the static and dynamic parts. For each case, we visualize the fitted OBBs together with the reconstructed 3D Gaussians [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] 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

41 extracted references · 11 linked inside Pith

  1. [1]

    In: Proceedings of the 3rd Conference on Robot Learning (2019)

    Abbatematteo, B., Tellex, S., Konidaris, G.: Learning to generalize kinematic mod- els to novel objects. In: Proceedings of the 3rd Conference on Robot Learning (2019)

    2019

  2. [2]

    arXiv preprint arXiv:1512.03012 (2015) 16 G

    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) 16 G. Lee et al

    Pith/arXiv arXiv 2015

  3. [3]

    Advances in Neural Information Processing Systems 35, 5982–5994 (2022)

    Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. Advances in Neural Information Processing Systems 35, 5982–5994 (2022)

    2022

  4. [4]

    Crc Press (2004)

    Ericson, C.: Real-time collision detection. Crc Press (2004)

    2004

  5. [5]

    Ester, M., Kriegel, H.P., Sander, J., Xu, X.: Density-based spatial clustering of ap- plications with noise. In: Int. Conf. knowledge discovery and data mining. vol. 240 (1996)

    1996

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via gen- eralizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081–7091 (2023)

    2023

  7. [7]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    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: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 27144– 27153 (2025)

    2025

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    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: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21201–21210 (2023)

    2023

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition

    Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 5616–5626 (2022)

    2022

  10. [10]

    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

  11. [11]

    arXiv preprint arXiv:2409.18121 (2024)

    Kerr, J., Kim, C.M., Wu, M., Yi, B., Wang, Q., Goldberg, K., Kanazawa, A.: Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. arXiv preprint arXiv:2409.18121 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    arXiv preprint arXiv:2508.02146 (2025)

    Kim, S., Ha, J., Kim, Y.H., Lee, Y., Park, F.C.: Screwsplat: An end-to-end method for articulated object recognition. arXiv preprint arXiv:2508.02146 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2017),https: //arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2017

  14. [14]

    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 (2023),https://arxiv.org/abs/2304.02643

    Pith/arXiv arXiv 2023

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, X., Wang, H., Yi, L., Guibas, L.J., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3706–3715 (2020)

    2020

  16. [16]

    arXiv preprint arXiv:2506.03594 (2025)

    Lin, S., Fang, J., Irshad, M.Z., Guizilini, V.C., Ambrus, R.A., Shakhnarovich, G., Walter, M.R.: Splart: Articulation estimation and part-level reconstruction with 3d gaussian splatting. arXiv preprint arXiv:2506.03594 (2025)

    Pith/arXiv arXiv 2025

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21136–21145 (2024)

    2024

  18. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Liu, J., Mahdavi-Amiri, A., Savva, M.: Paris: Part-level reconstruction and motion analysis for articulated objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 352–363 (2023)

    2023

  19. [19]

    arXiv preprint arXiv:2302.14268 (2023) StructureGS 17

    Liu, X., Zhang, J., Hu, R., Huang, H., Wang, H., Yi, L.: Self-supervised category- level articulated object pose estimation with part-level se (3) equivariance. arXiv preprint arXiv:2302.14268 (2023) StructureGS 17

    Pith/arXiv arXiv 2023

  20. [20]

    arXiv preprint arXiv:2502.19459 (2025)

    Liu, Y., Jia, B., Lu, R., Ni, J., Zhu, S.C., Huang, S.: Artgs: Building interactable replicas of complex articulated objects via gaussian splatting. arXiv preprint arXiv:2502.19459 (2025)

    Pith/arXiv arXiv 2025

  21. [21]

    In: 2024 International Conference on 3D Vision (3DV)

    Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800–809. IEEE (2024)

    2024

  22. [22]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460–4470 (2019)

    2019

  23. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Mu, J., Qiu, W., Kortylewski, A., Yuille, A., Vasconcelos, N., Wang, X.: A-sdf: Learning disentangled signed distance functions for articulated shape representa- tion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13001–13011 (2021)

    2021

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Noguchi, A., Iqbal, U., Tremblay, J., Harada, T., Gallo, O.: Watch it move: Unsu- pervised discovery of 3d joints for re-posing of articulated objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3677–3687 (2022)

    2022

  25. [25]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without su...

    Pith/arXiv arXiv 2024

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Park,J.J.,Florence,P.,Straub,J.,Newcombe,R.,Lovegrove,S.:Deepsdf:Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165– 174 (2019)

    2019

  27. [27]

    Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)

    2017

  28. [28]

    arXiv preprint arXiv:2502.02590 (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. arXiv preprint arXiv:2502.02590 (2025)

    Pith/arXiv arXiv 2025

  29. [29]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    Schönberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    2016

  30. [30]

    In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06)

    Seitz, S.M., Curless, B., Diebel, J., Scharstein, D., Szeliski, R.: A comparison and evaluation of multi-view stereo reconstruction algorithms. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06). vol. 1, pp. 519–528. IEEE (2006)

    2006

  31. [31]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettle- moyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10740–10749 (2020)

    2020

  32. [32]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Song, C., Wei, J., Foo, C.S., Lin, G., Liu, F.: Reacto: Reconstructing articulated objects from a single video. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 5384–5395 (2024)

    2024

  33. [33]

    Advances in neural information processing systems34, 251–266 (2021) 18 G

    Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N.,Mukadam,M.,Chaplot,D.S.,Maksymets,O.,etal.:Habitat2.0:Traininghome assistants to rearrange their habitat. Advances in neural information processing systems34, 251–266 (2021) 18 G. Lee et al

    2021

  34. [34]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025)

    2025

  35. [35]

    In: Proceedings of the European conference on computer vision (ECCV)

    Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52–67 (2018)

    2018

  36. [36]

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π 3: Permutation-equivariant visual geometry learning (2026), https://arxiv.org/abs/2507.13347

    Pith/arXiv arXiv 2026

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wei, F., Chabra, R., Ma, L., Lassner, C., Zollhöfer, M., Rusinkiewicz, S., Sweeney, C., Newcombe, R., Slavcheva, M.: Self-supervised neural articulated shape and appearance models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15816–15826 (2022)

    2022

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    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: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3141–3150 (2024)

    2024

  39. [39]

    ACM Transactions on Graphics (TOG) 28(3), 1–9 (2009)

    Xu, W., Wang, J., Yin, K., Zhou, K., Van De Panne, M., Chen, F., Guo, B.: Joint- aware manipulation of deformable models. ACM Transactions on Graphics (TOG) 28(3), 1–9 (2009)

    2009

  40. [40]

    Ye, V., Li, R., Kerr, J., Turkulainen, M., Yi, B., Pan, Z., Seiskari, O., Ye, J., Hu, J., Tancik, M., Kanazawa, A.: gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research26(34), 1–17 (2025) StructureGS 19 7 Supplementary Material In this supplementary material, we provide additional details on implementation, robustnes...

    2025

  41. [41]

    Point cloud noise

    The articulation transformsT k 1 are initialized to identity for all parts. The initialization is intentionally coarse. Its role is to place the oriented boxes near plausible object parts, rather than to provide accurate segmentation. During optimization, the Gaussian primitives, OBB parameters, and articulation parameters are jointly refined by photometr...

    arXiv 2029

This paper was first reviewed by grok-4.5 on July 30, 2026.