Pith. sign in

REVIEW 3 major objections 5 minor 62 references

DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read DIPO shows that two images of an object—one at rest and one articulated—give a diffusion model enough motion information to generate accurate part layouts, joint parameters, and connectivity graphs for articulated 3D objects, and that a…

desk verdict Solid dual-image articulated generation method and dataset, but the PM-X retrieval split could leak test geometry and needs explicit clarification. read the letter →

arxiv 2505.20460 v2 pith:OIIWX5VT submitted 2025-05-26 cs.CV

classification cs.CV
keywords articulated3Dobjectgenerationdual-stateimageconditioningdiffusionmodelpartconnectivitygraphchain-of-thoughtreasoningsyntheticdatasetPM-XNet-Mobilitycontrollable
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

This paper argues that conditioning articulated 3D object generation on two images—one of the object at rest and one after it has moved—gives a generative model the motion information needed to predict where parts sit, how they hinge or slide, and which parts connect to which. The proposed system, DIPO, combines a diffusion transformer with a chain-of-thought graph reasoner: the image pair is fused through cross-attention, and the predicted connectivity graph is fed back as an attention mask during denoising. The paper also introduces PM-X, a large synthetic dataset of structurally complex articulated objects built by an automated LLM-based pipeline, and shows that training on it improves generalization to out-of-distribution objects. The practical motivation is that capturing two photos of an object is cheap, and success would make it much easier to build interactive 3D scenes for robotics and simulation.

What carries the argument

The load-bearing mechanism is the dual-state injection: DINOv2 features of the resting and articulated images are combined through two cross-attention steps so that the articulated features first attend to the resting features and the result is injected into the denoising transformer. A separate chain-of-thought graph reasoner detects candidate parts, applies structural rules, and outputs an adjacency matrix that masks self-attention, making the connectivity prediction explicit. On the data side, the LEGO-Art pipeline assembles part primitives from PartNet-Mobility into grid-layout objects using LLM agents, converts grid coordinates to URDF joint parameters, renders dual-state image pairs, and filters implausible samples with a VLM.

What would settle it

Compare DIPO's graph accuracy on real-world pairs where the articulated image is a genuine physical capture versus pairs where it was synthesized by GPT-4o; a large gap would show the method depends on the quality of the synthesized image, not on genuine motion evidence. Also, feeding a pair of images of two different objects of the same category should produce an incoherent connectivity graph if the model truly uses motion cues.

Watch

Extended reading notes

Core claim

DIPO establishes that conditioning a diffusion transformer on a pair of images—one resting, one articulated—is enough to generate an articulated 3D object with accurate part layouts, joint parameters, and part connectivity. In the paper's experiments this dual-image input outperforms single-image baselines (URDFormer, NAP-ICA, SINGAPO) on both resting and articulated metrics on the PartNet-Mobility test set and on the out-of-distribution ACD set, and it shows a smaller drop in accuracy when moving from the resting to the articulated state, indicating that the second image supplies genuine motion cues. The accompanying PM-X dataset, 600 objects with an average of 19.4 parts, is shown by ablations to be responsible for much of the generalization to complex structures.

Load-bearing premise

The approach assumes that a valid, accurate second photo showing the object in motion is available or can be generated for every input; if that second photo is wrong, the method loses its edge over single-image models.

Editorial extensions

If this is right

  • Dual-state image pairs become a practical interface for interactive 3D content creation: photograph an object closed and open, get an animatable model.
  • The smaller resting-to-articulated performance gap means articulated-state rendering can be controlled more reliably, which matters for robot manipulation planning.
  • The PM-X dataset and LEGO-Art pipeline can be extended beyond storage furniture and tables to other categories, serving as a scalable source of training data for articulation reasoning.
  • CoT-based graph reasoning with visual prompts may generalize to other structured-prediction tasks where pairwise relationships must be inferred from images.

Reading between the lines

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

  • Inference: The same dual-image conditioning could be applied to deformable objects such as clothing or cables, where motion cues reveal non-rigid part relationships, though the paper's graph representation would need extension.
  • Inference: The paper's reliance on GPT-4o-synthesized articulated images for real-world inputs is untested; a user study or a captured-pair benchmark would clarify whether DIPO truly works in the wild or only on synthetic or consistent pairs.
  • Inference: LEGO-Art's text-to-grid-to-URDF pipeline could double as a text-to-articulated-object generator, which the paper notes but does not evaluate.
  • Inference: Since the graph reasoner and the diffusion model are trained jointly but the graph is predicted before denoising, errors in the predicted graph constrain the generation; measuring end-to-end sensitivity to graph errors would be informative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes DIPO, a diffusion-transformer framework that takes a resting-state and an articulated-state image pair of an object and generates part layouts, joint parameters, and a part connectivity graph for an articulated 3D object. It also introduces LEGO-Art, an LLM-agent pipeline that assembles part meshes from PartNet-Mobility into a new 600-object PM-X dataset, and uses GPT-4o-generated visual prompts for a CoT-based Graph Reasoner. Experiments compare against URDFormer, NAP-ICA, and SINGAPO on PM and ACD test sets, with ablations over the PM-X data, the dual-state injection module, and the graph reasoner.

Significance. If the reported results hold, DIPO would be a useful step toward controllable articulated-object generation from easy-to-acquire dual-state images, and PM-X could become a valuable resource for training and evaluating such models. Strengths include the clear task formulation, the fully automated dataset pipeline with URDF and rendered-image outputs, the explicit ablation of each component, and the stated intention to release code and data. The main weaknesses are the unresolved possibility of geometry leakage from the test set into PM-X, the confounded comparison with baselines that do not appear to train on PM-X, and the absence of variance or significance reporting.

major comments (3)
  1. [§4.1, §5.1] PM-X is built by retrieving part meshes from PartNet-Mobility, and Section 5.1 trains on 493 PM objects plus 600 PM-X samples while testing on 77 held-out PM objects; since PM contains 570 objects in total, the paper must state explicitly that the retrieval pool excludes the 77 test objects. Without such an exclusion, held-out test geometry can appear in PM-X training via exact or near-exact mesh retrieval, so the PM metrics in Table 2 could be inflated by memorization rather than generalization. Please clarify the retrieval pool and, if it was not restricted, re-run the PM evaluation with a leakage-free split.
  2. [§5.1, §5.2.1, Tables 2–3] The headline comparison is confounded by training data: DIPO is trained on 493 PM objects plus 600 PM-X samples, while the retrained SINGAPO and finetuned URDFormer are not described as receiving PM-X. Since the paper's own ablation shows PM-X improves performance, the advantage in Tables 2 and 3 may be attributable to the extra data rather than the proposed architecture. A clean comparison should include a DIPO variant trained without PM-X on the same PM test set, or train the baselines on the same combined data.
  3. [§5.2.2] The paper reports metric values averaged over five stochastic samples per test input, but no standard deviations, confidence intervals, or per-sample spread are given. Given the abstract's claim of 'significantly outperforms,' the absence of statistical reporting makes it impossible to judge whether the differences in Tables 2 and 3 are meaningful, particularly where margins are small. Please report variance or run significance tests.
minor comments (5)
  1. [§5.3, Figure 6] The text says 'IoU scores ... degrade steadily as the PM-X ratio increases,' which contradicts the claim that PM-X helps; this should read 'as the ratio decreases' (or the figure should be rechecked).
  2. [§5.1] The paper says 77 held-out objects are rendered from two random views, producing 144 dual-state test samples; 77 × 2 = 154, so this number should be corrected or explained.
  3. [§5.2.3] For Internet-collected real-world examples, the articulated-state image is generated by GPT-4o and is not independently verified; this should be stated as a limitation of the qualitative real-world evaluation.
  4. [Abstract / throughout] There are numerous typos and grammatical slips (e.g., 'issuse,' 'futher,' 'rest ing,' 'precdiction,' 'fisrt'); the manuscript would benefit from a careful proofreading pass.
  5. [§3.3, Appendix B] The Graph Reasoner is described as few-shot learning on GPT-4o-synthesized visual prompts, but the number of prompts, their filtering criteria, and how they are used for few-shot learning are not specified, making this component difficult to reproduce.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: DIPO's empirical claims rest on external PM/ACD benchmarks; self-citations are minor, while the PM-X/PM retrieval overlap is a data-contamination risk rather than a demonstrated circular reduction.

full rationale

The paper's central claims are empirical rather than derivational: DIPO is a trained diffusion model whose outputs are compared against held-out ground-truth annotations from PartNet-Mobility and ACD, so the reported predictions are not equal by construction to any fitted input. The dual-state conditioning module in Eq. (1) is an architectural mechanism, not a definition of the evaluation metrics, and the Graph Reasoner is trained on GPT-4o visual prompts but tested on real graph labels. The only self-citations (e.g., LAMP [45] and AR-1-to-3 [51]) appear in related-work context and do not carry the argument. One caveat is data hygiene: Section 4.1 states that PM-X geometry is built by 'retrieving mesh primitives from PartNet-Mobility,' and Section 5.1 adds 600 PM-X samples to the training set while evaluating on 77 'held-out' PM objects; the paper never states that the retrieval pool excludes the held-out test objects, so Table 2 could be inflated by geometry leakage if the pool includes test meshes. Because the paper neither confirms nor denies the exclusion, this is a missing-support concern rather than an exhibited circular reduction, and Table 3 and Table 5 on the external ACD benchmark independently support the generalization claim.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or fitted constants. Its central assumptions are about input informativeness, data annotation quality, and the validity of the synthetic data generation pipeline.

assumptions (3)
  • domain assumption A dual-state image pair encodes sufficient motion cues and connectivity information to infer articulation.
    Section 1 states that the dual-state image pair 'encodes essential motion cues and connectivity information.' The entire method and its claimed advantage rely on this premise.
  • domain assumption Part-level ground-truth annotations in PM, ACD, and PM-X are accurate enough to supervise and evaluate the model.
    Training and all quantitative metrics (dgIoU, dcDist, dCD, Acc) depend on the correctness of part bounding boxes, joint parameters, and connectivity graphs in the benchmark datasets.
  • domain assumption Retrieval-based assembly from PartNet-Mobility primitives produces physically plausible articulated objects after VLM filtering.
    Section 4.1 describes a pipeline that recombines part primitives and uses a VLM filter to keep plausible samples. The usefulness of PM-X for training and generalization depends on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data." pith.science (2026). https://pith.science/paper/OIIWX5VT

@misc{pith2026250520460,
  author       = {Pith},
  title        = {Pith review of: DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIIWX5VT}},
  note         = {Machine review of arXiv:2505.20460}
}
read the original abstract

We present DIPO, a novel framework for the controllable generation of articulated 3D objects from a pair of images: one depicting the object in a resting state and the other in an articulated state. Compared to the single-image approach, our dual-image input imposes only a modest overhead for data collection, but at the same time provides important motion information, which is a reliable guide for predicting kinematic relationships between parts. Specifically, we propose a dual-image diffusion model that captures relationships between the image pair to generate part layouts and joint parameters. In addition, we introduce a Chain-of-Thought (CoT) based graph reasoner that explicitly infers part connectivity relationships. To further improve robustness and generalization on complex articulated objects, we develop a fully automated dataset expansion pipeline, name LEGO-Art, that enriches the diversity and complexity of PartNet-Mobility dataset. We propose PM-X, a large-scale dataset of complex articulated 3D objects, accompanied by rendered images, URDF annotations, and textual descriptions. Extensive experiments demonstrate that DIPO significantly outperforms existing baselines in both the resting state and the articulated state, while the proposed PM-X dataset further enhances generalization to diverse and structurally complex articulated objects. Our code and dataset will be released to the community upon publication.

Figures

Figures reproduced from arXiv: 2505.20460 by the authors.

Figure 1
Figure 1. Visual comparison on real-captured data. (a) SINGAPO struggles with challenging data and fails to model motion relationships due to its reliance on a single input. However, our DIPO (b), which conditioned on dual-state image pairs, effectively generates accurate layouts and enables precise control if part motion across different articulated states. single RGB image, to synthesize articulated objects directly. Howeve… view at source ↗
Figure 2
Figure 2. Overview of the proposed DIPO framework. The left part shows the proposed LEGO-Art [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Dual-state visual prompt used by the Graph Reasoner. GPT-4o can produce realistic and structurally complex image pairs. We condition the denoising process on both resting￾state and articulated-state images to capture motion￾aware cues. Let FR and FA denote the DI￾NOv2 [29] features from the resting and articulated images, respectively. To integrate these into the dif￾fusion network, we apply a Dual-State Injection M… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: An overview of the fully automated synthesis pipeline for the proposed PM-X dataset. The [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison between the proposed DIPO and two baselines. The fist two columns [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablative comparison under different ratios of PM-X data using only 25% and 50% of the PM-X data. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The system prompt of Descrition Roller. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: The system prompt of Layout Builder. This agent is inspired by the code of scripting toolkit [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: The system prompt of Visual Filter. B Details of Graph Reasoner The proposed Graph Reasoner can infer articulated connectivity from a dual-state image pair based on chain-of-thought [41, 17] prompt, which is illustrated as followed: Graph Reasoner: You are an expert in…
Figure 10
Figure 10. Figure 10: The system prompt of Graph Reasoner. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Each pair shows a closed-state image (left) and an open-state image (right) of an articulated [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: More visual examples of PM-X dataset. Each example includes: (1) the part layout [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 38 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

  2. [2]

    Parametric correspondence and chamfer matching: Two new techniques for image matching

    Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In Proceedings: Image Understanding Workshop, pages 21–27. Science Applications, Inc, 1977

  3. [3]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems , 33:1877–1901, 2020

  4. [4]

    Shapenet: An information-rich 3d model repository

    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. [5]

    Fast-snarf: A fast deformer for articulated neural fields

    Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11796–11809, 2023

  6. [6]

    Urdformer: A pipeline for constructing articulated simulation environments from real-world images

    Zoey Chen, Aaron Walsman, Marius Memmel, Kaichun Mo, Alex Fang, Karthikeya Vemuri, Alan Wu, Dieter Fox, and Abhishek Gupta. Urdformer: A pipeline for constructing articulated simulation environments from real-world images. arXiv preprint arXiv:2405.11656, 2024

  7. [7]

    Abo: Dataset and benchmarks for real-world 3d object understanding

    Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126–21136, 2022

  8. [8]

    3d-future: 3d furniture shape with texture

    Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision , 129:3313–3337, 2021

Show all 62 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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15752–15761, 2021

  2. [10]

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

  3. [11]

    Active articulation model estimation through interactive perception

    Karol Hausman, Scott Niekum, Sarah Osentoski, and Gaurav S Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3305–3312. IEEE, 2015

  4. [12]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

  5. [13]

    Learning to predict part mobility from a single static snapshot

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

  6. [14]

    S2o: Static to openable enhancement for articulated 3d objects

    Denys Iliash, Hanxiao Jiang, Yiming Zhang, Manolis Savva, and Angel X Chang. S2o: Static to openable enhancement for articulated 3d objects. arXiv preprint arXiv:2409.18896, 2024

  7. [15]

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation

    Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigat...

  8. [16]

    Parahome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions

    Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. Parahome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 10

  9. [17]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems , 35:22199–22213, 2022

  10. [18]

    Nap: Neural 3d articulated object prior

    Jiahui Lei, Congyue Deng, William B Shen, Leonidas J Guibas, and Kostas Daniilidis. Nap: Neural 3d articulated object prior. Advances in Neural Information Processing Systems , 36:31878–31894, 2023

  11. [19]

    igibson 2.0: Object-centric simulation for robot learning of everyday household tasks

    Chengshu Li, Fei Xia, Roberto Martín-Martín, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021

  12. [20]

    Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation

    Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martín-Martín, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Conference on ...

  13. [21]

    Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling

    Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 19711–19722, 2024

  14. [22]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 352–363, 2023

  15. [23]

    Singapo: Single image controlled generation of articulated parts in objects

    Jiayi Liu, Denys Iliash, Angel X Chang, Manolis Savva, and Ali Mahdavi-Amiri. Singapo: Single image controlled generation of articulated parts in objects. arXiv preprint arXiv:2410.16499, 2024

  16. [24]

    Cage: controllable articulation generation

    Jiayi Liu, Hou In Ivan Tam, Ali Mahdavi-Amiri, and Manolis Savva. Cage: controllable articulation generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 17880–17889, 2024

  17. [25]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38–55. Springer, 2024

  18. [26]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

  19. [27]

    Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

  20. [28]

    Where2act: From pixels to actions for articulated 3d objects

    Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813–6823, 2021

  21. [29]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

  22. [30]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4195–4205, 2023

  23. [31]

    Habitat 3.0: A co-habitat for humans, avatars and robots

    Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023

  24. [32]

    Understanding 3d object interaction from a single image

    Shengyi Qian and David F Fouhey. Understanding 3d object interaction from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision , pages 21753–21763, 2023

  25. [33]

    3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting

    Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5020–5030, 2024

  26. [34]

    O’Reilly Media, Inc

    Morgan Quigley, Brian Gerkey, and William D Smart. Programming Robots with ROS: a practical introduction to the Robot Operating System . " O’Reilly Media, Inc.", 2015. 11

  27. [35]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

  28. [36]

    igibson 1.0: A simulation environment for interactive tasks in large realistic scenes

    Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez- D’Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. igibson 1.0: A simulation environment for interactive tasks in large realistic scenes. In 2021 IEEE/RSJ Internation...

  29. [37]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

  30. [38]

    3d warehouse, 2025

    Trimble Inc. 3d warehouse, 2025. Accessed: 2025-05-14

  31. [39]

    Cla-nerf: Category-level articulated neural radiance field

    Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. Cla-nerf: Category-level articulated neural radiance field. In 2022 International Conference on Robotics and Automation (ICRA) , pages 8454–8460. IEEE, 2022

  32. [40]

    Shape2motion: Joint analysis of motion parts and attributes from 3d shapes

    Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2motion: Joint analysis of motion parts and attributes from 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8876–8884, 2019

  33. [41]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

  34. [42]

    Photo wake-up: 3d character animation from a single photo

    Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908–5917, 2019

  35. [43]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

  36. [44]

    Model soups: aver- aging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: aver- aging weights of multiple fine-tuned models improves accuracy without increasing in...

  37. [45]

    Lamp: Learn a motion pattern for few-shot video generation

    Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7089–7098, 2024

  38. [46]

    Sapien: A simulated part-based interactive environment

    Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11...

  39. [47]

    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

  40. [48]

    Banmo: Building animatable 3d neural models from many casual videos

    Gengshan Yang, Minh V o, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863–2873, 2022

  41. [49]

    Object wake-up: 3d object rigging from a single image

    Ji Yang, Xinxin Zuo, Sen Wang, Zhenbo Yu, Xingyu Li, Bingbing Ni, Minglun Gong, and Li Cheng. Object wake-up: 3d object rigging from a single image. In European Conference on Computer Vision , pages 311–327. Springer, 2022

  42. [50]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3836–3847, 2023

  43. [51]

    tray" parts are only allowed if the object is a microwave. Each object must contain exactly one implicit

    Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. 12 Abstract Our supplementary materi...

  44. [52]

    A closed-state image (all movable parts in their fully closed positions)

  45. [53]

    Workflow:

    An open-state image (all movable parts in their fully opened positions) Your task is to analyze the object’s articulated structure and generate a connectivity graph describing the part relationships. Workflow:

  46. [54]

    base"; "tray

    Part Detection • Detect candidate parts in the closed-state image, optionally using the open-state image to resolve ambiguity or occlusion. • Allowed part types: [’base’, ’door’, ’drawer’, ’handle’, ’knob’, ’tray’] • Ignore small decorative elements attached directly to the ba...

  47. [55]

    drawer_1 (attached to base)

    Step-by-Step Reasoning (a) Part Listing: List all detected parts and their counts (no attachment inference yet). (b) Validation: Enforce structural rules: • Exactly one base • Each door or drawer may have at most two handles or knobs • Every handle/knob must be attached to a d...

  48. [56]

    Randomly replace small parts like handles and knobs with those of other objects, and perturb their positions

  49. [57]

    Randomly rescale the whole object

  50. [58]

    Rotate the whole object upside-down

  51. [59]

    The other one focus on joint-level augmentation:

    Stacking several objects together to build more complex objects. The other one focus on joint-level augmentation:

  52. [60]

    Change the revolute joint into prismatic joint. 16

  53. [61]

    Randomly modify the direction of revolute joint

  54. [62]

    E Limitations & Future Work We follow the experimental settings of SINGAPO [23] for a fair comparison

    Randomly fix the joint. E Limitations & Future Work We follow the experimental settings of SINGAPO [23] for a fair comparison. However, the benchmark used in SINGAPO only contains several categories, which especially focuses on cabinet-like objects. This limited object diversi...

Pith tools

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