MonoArt: Progressive Structural Reasoning for Monocular Articulated 3D Reconstruction

Haitian Li; Haozhe Xie; Junxiang Xu; Beichen Wen; Fangzhou Hong; Ziwei Liu

cs.CVMar 19, 2026

MonoArt: Progressive Structural Reasoning for Monocular Articulated 3D Reconstruction

Authors: Haitian Li, Haozhe Xie, Junxiang Xu, Beichen Wen, Fangzhou Hong, Ziwei Liu

Organizations: S-Lab, Nanyang Technological University, 637335 Singapore

Abstract

Reconstructing articulated 3D objects from a single image requires jointly inferring object geometry, part structure, and motion parameters from limited visual evidence. A key difficulty lies in the entanglement between motion cues and object structure, which makes direct articulation regression unstable. Existing methods address this challenge through multi-view supervision, retrieval-based assembly, or auxiliary video generation, often sacrificing scalability or efficiency. We present MonoArt, a unified framework grounded in progressive structural reasoning. Rather than predicting articulation directly from image features, MonoArt progressively transforms visual observations into canonical geometry, structured part representations, and motion-aware embeddings within a single architecture. This structured reasoning process enables stable and interpretable articulation inference without external motion templates or multi-stage pipelines. Extensive experiments on PartNet-Mobility demonstrate that OM achieves state-of-the-art performance in both reconstruction accuracy and inference speed. The framework further generalizes to robotic manipulation and articulated scene reconstruction.

Figures & tables

Figure 1: (Left) Qualitative results of SINGAPO [ 24 ] , Articulate-Anything (ArtAny) [ 18 ] , PhysX-Anything (PhysXAny) [ 4 ] , and MonoArt on diverse objects. (Right) F-score vs. inference time on the PartNet-Mobility [ 51 ] test set. Circles indicate models evaluated on 7 categories, while triangles denote models supporting all 46 categories.

Figure 2: Overview of MonoArt. TRELLIS-based 3D Generator reconstructs a canonical shape from a single image. Part-Aware Semantic Reasoner derives tri-plane-based part embeddings. Dual-Query Motion Decoder performs iterative motion reasoning, and Kinematic Estimator predicts part-level articulation parameters (motion type, origin, axis, limits) and infers the kinematic tree structure. Note that “Attn.”, “Interp.”, “Proj.”, “Cont.”, “Trans.”, and “Init.” represent “Attention”, “Interpolation”, “Projection”, “Contrast”, “Transformer”, and “Initialization”, respectively. $\oplus$ and $\otimes$ denote element-wise addition and matrix multiplication, respectively.

Partial 7 classes
Method	Geometry				Kinematics
Method	CD $\downarrow$	F-Score $\uparrow$	PSNR $\uparrow$	CLIP $\uparrow$	Type Acc. $\uparrow$	Axis Err. $\downarrow$	Pivot Err. $\downarrow$
URDFormer [ 8 ]	4.73	0.275	12.43	0.845	35.22	1.324	0.404
SINGAPO [ 24 ]	1.26	0.572	15.22	0.870	77.12	0.493	0.201
MonoArt	0.77	0.728	17.55	0.926	88.26	0.209	0.085
All 46 classes

Table 1: Quantitative comparison on the PartNet-Mobility dataset. CD is scaled by

\times 10^{-2}

. Type Acc. (%) denotes joint type classification accuracy. Axis Err. (rad) represents joint axis direction error.. Pivot Err. is joint pivot distance error in normalized object coordinates. Best results are highlighted in bold.

Figure 3: Qualitative results on the test set of PartNet-Mobility. ArtAny and PhysXAny denote Articulate-Anything and PhysXAnything, respectively. For each object, we show the reconstructed geometry under two sampled articulated states.

Figure 4: Qualitative results on in-the-wild images. ArtAny and PhysXAny denote Articulate-Anything and PhysXAnything, respectively. For each object, we show the reconstructed geometry under two sampled articulated states.

Enabled	Loss	Geometry			Kinematics
	Loss	CD $\downarrow$	F-Score $\uparrow$	PSNR $\uparrow$	Type Acc. $\uparrow$	Axis Err. $\downarrow$	Pivot Err. $\downarrow$
✗	✗	1.74	0.626	17.96	24.72	0.549	0.237
✓	✗	1.63	0.643	17.74	41.60	0.922	0.323
✓	CE	1.49	0.648	17.71	57.74	1.029	0.302
✓	Tri.	1.25	0.670	18.55	67.47	0.423	0.108

Table 2: Ablation of Part-Aware Semantic Reasoner. CD is scaled by

\times 10^{-2}

. Type Acc. (%) is joint classification accuracy. Axis Err. (rad) and Pivot Err. denote axis direction and pivot distance errors. Note that “CE” and “Tri.” denote “Cross-Entropy” and “Triplet”, respectively.

DQI.	Res.	$L$	Geometry			Kinematics
			CD $\downarrow$	F-Score $\uparrow$	PSNR $\uparrow$	Type Acc. $\uparrow$	Axis Err. $\downarrow$	Pivot Err. $\downarrow$
✗	Both	6	1.67	0.622	18.11	44.06	0.472	0.329
✓	$Q_{p}^{l}$	6	1.73	0.640	17.81	66.41	0.523	0.181
✓	$Q_{c}^{l}$	6	1.29	0.663	17.80	60.88	0.506	0.184
✓	Both	0	1.70	0.652	17.94	62.65	0.640	0.186
✓	Both	1	1.71	0.652	17.91	63.12	0.608	0.189

Table 3: Ablation of Dual-Query Motion Decoder. CD is scaled by

\times 10^{-2}

. Type Acc. (%) is joint classification accuracy. Axis Err. (rad) and Pivot Err. denote axis direction and pivot distance errors. Note that “DQI.” indicates whether Dual-Query Initialization is enabled. “Res.” denotes residual updates in the position and/or content query branches.

L

is the number of refinement blocks.

$\mathbf{Q}_{p}^{L}$	$\mathbf{H}$	Geometry			Kinematics
		CD $\downarrow$	F-Score $\uparrow$	PSNR $\uparrow$	Type Acc. $\uparrow$	Axis Err. $\downarrow$	Pivot Err. $\downarrow$
✗	✓	1.65	0.652	17.92	67.20	0.499	0.191
✓	✗	2.35	0.573	18.11	27.14	0.882	0.283
✓	✓	1.25	0.670	18.55	67.47	0.423	0.108

Table 4: Ablation of Kinematic Estimator. CD is scaled by

\times 10^{-2}

. Type Acc. (%) is joint classification accuracy. Axis Err. (rad) and Pivot Err. denote axis direction and pivot distance errors.

Figure 5: Robot manipulation with generated articulated objects. MonoArt reconstructions are directly imported into IsaacSim for contact-rich interaction.

Figure 6: Articulated scene reconstruction. MonoArt augments SAM3D [ 42 ] with object-level articulation recovery to produce articulated, operable 3D scenes.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Visualization of point features $\mathbf{H}$ from the Part-Aware Semantic Reasoner. Points are colored by PCA projection of the features. “Enabled” indicates whether the reasoner is used, and “Loss” specifies the supervision type. Note that “CE” and “Tri.” denote “Cross-Entropy” and “Triplet”, respectively.

Figure 8: Representative Failure Cases. (Top) an object containing an extremely small component (printer button). (Bottom) an unseen articulated object (bicycle).

Articulated Object Reconstruction

Articulated object reconstruction recovers 3D part geometry, joints, and motion from images, video, or depth for digital twins and robot interaction.

See all 60 papers

Track2Art: Articulated Object Model Recovery with Visual-Geometric Track Representations
Sep 23, 2026Xiaotong Li, Yixiong Jing, Junsheng Ding +4Articulated Object Reconstruction 3D Part Segmentation
Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K. · School of Engineering and Design, Technical University of Munich · Munich Center for Machine Learning (MCML), Munich, Germany.
Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning
Jun 20, 2026Xuyang Wang, Zhenyu Li, Jian Ding +4Articulated Object Reconstruction 3D Reconstruction
Australian National University, Canberra, Australia · King Abdullah University of Science and Technology, Thuwal, Saudi Arabia
Articulat3D: Reconstructing Articulated Digital Twins From Monocular Videos with Geometric and Motion Constraints
Mar 12, 2026Lijun Guo, Haoyu Zhao, Xingyue Zhao +5Articulated Object Reconstruction 3D Reconstruction
Wuhan University · Hong Kong Embodied AI Lab · The Chinese University of Hong Kong +3

Sep 23, 2026J/K move · Enter open · S save

Track2Art: Articulated Object Model Recovery with Visual-Geometric Track Representations

Xiaotong Li, Yixiong Jing, Junsheng Ding, Weihang Li+3

Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K. · School of Engineering and Design, Technical University of Munich · Munich Center for Machine Learning (MCML), Munich, Germany.

Articulated Object Reconstruction 3D Part Segmentation

Open paper AskPDF arXiv

Understanding articulated objects is fundamental for robotic interaction, requiring accurate rigid-part discovery and the recovery of their kinematic relations. Existing approaches often treat articulation as a by-product of reconstructed geometry or recover it through per-instance optimization. We instead build on the hypothesis that articulation is directly observable from persistent motion: points on the same rigid part move coherently, while relative motion between parts reveals their kinematic constraints. We present Track2Art, a motion-centric framework for recovering structured articulated objects from RGB-D interaction videos. Track2Art lifts tracked image points into persistent 3D trajectories and combines pretrained tracking features, visual descriptors, and explicit trajectory geometry. These representations are grouped into a variable number of rigid-part hypotheses and subsequently used to recover directed kinematic relations, joint types, and joint geometry through rotation-equivariant learned--analytic reasoning. On the aligned 20-object PartNet-Mobility suite, Track2Art achieves 0.695 Point IoU and 0.410 end-to-end J@20, while requiring neither ground-truth part counts nor test-time optimization.

Abstract

Articulated Object Reconstruction

Explore similar work

Track2Art: Articulated Object Model Recovery with Visual-Geometric Track Representations

Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning

Articulat3D: Reconstructing Articulated Digital Twins From Monocular Videos with Geometric and Motion Constraints