pith. machine review for the scientific record. sign in

arxiv: 2603.28178 · v2 · submitted 2026-03-30 · 💻 cs.CV

Recognition: unknown

ToLL: Topological Layout Learning with Asymmetric Cross-View Structural Distillation for 3D Scene Graph Generation Pretraining

Yucheng Huang , Luping Ji , Xiangwei Jiang , Wen Li , Mao Ye
Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:56 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D scene graphpretrainingtopological learninggeometric shortcutself-distillationgraph neural networkspatial reasoning
0
0 comments X

The pith

ToLL prevents geometric shortcuts in 3D scene graph pretraining by forcing topological learning via an information bottleneck.

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

This paper introduces Topological Layout Learning (ToLL) to fix the geometric shortcut problem in pretraining 3D scene graph models. Generative pretraining often lets models reconstruct scenes by interpolating dense object positions rather than learning the topological constraints encoded in predicate edges. ToLL uses anchor-conditioned topological geometry reasoning on zero-centered subgraphs with a recurrent GNN and one sparse anchor to create an information bottleneck. This compels recovery of layouts using predicate representations, paired with structural multi-view augmentation for self-distillation to preserve semantics. Sympathetic readers would care because it could produce better generalizable representations for 3D spatial understanding without relying on dense geometric cues.

Core claim

The central discovery is that Anchor-Conditioned Topological Geometry Reasoning recovers the global layout of zero-centered subgraphs using a recurrent GNN conditioned on a single anchor with sparse spatial prior, creating an information bottleneck that forces the model to learn from predicate representations rather than geometric interpolation. Structural Multi-view Augmentation then enables asymmetric cross-view structural distillation to enhance representations without semantic corruption.

What carries the argument

Anchor-Conditioned Topological Geometry Reasoning, a recurrent GNN mechanism that reconstructs layouts from zero-centered subgraphs with one anchor to enforce predicate-based learning.

If this is right

  • Pretrained 3DSG models outperform state-of-the-art baselines on generation tasks.
  • Models prioritize topological constraints over spatial priors in reconstruction.
  • Self-distillation improves representation quality while avoiding semantic corruption from augmentations.
  • Better pretraining supports improved spatial understanding and affordance perception.

Where Pith is reading between the lines

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

  • This bottleneck technique might apply to other pretraining settings where geometric features overshadow relational learning.
  • It could enhance performance in downstream 3D tasks like navigation or object interaction with fewer labels.
  • Varying the number of anchors or subgraph sizes offers a testable way to tune the information bottleneck strength.

Load-bearing premise

The setup with zero-centered subgraphs and a single sparse anchor truly blocks geometric interpolation, leaving no choice but to use predicate representations for layout recovery.

What would settle it

A model using ToLL that still reconstructs scenes accurately by interpolating positions alone, without needing predicates, on novel test layouts would disprove the claim.

Figures

Figures reproduced from arXiv: 2603.28178 by Luping Ji, Mao Ye, Wen Li, Xiangwei Jiang, Yucheng Huang.

Figure 1
Figure 1. Figure 1: Proxy task scheme comparison of 3DSG generation: (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The motivation of designing Anchor-Conditioned Topological Geometric Reasoning (ACTGR). We observe that exposing [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Our 3D Scene Graph Pretraining scheme via Topological Layout Learning with Structural Multi-view Augmentation. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Predicate A@1 for all predicate categories. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicate A@1 for all predicate categories. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Analysis of accuracy curves (Train and Validation). [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of latent features clustering. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of clustering confusion matrix. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
read the original abstract

3D Scene Graph (3DSG) generation plays a pivotal role in spatial understanding and affordance perception. To mitigate generalization issues from data scarcity, joint-embedding and generative proxy tasks are proposed to pre-train 3DSG representations on predicate label-free datasets. Currently, generative pre-training usually bypasses the semantic corruption caused by the geometric augmentations in joint-embedding, but cannot avoid a negative problem ``Geometric Shortcut." In this problem, exposing dense object spatial and scale priors will induce models to trivially reconstruct scenes by interpolating object positions, rather than learning the underlying topological constraints provided by edges. To address this issue, we propose a Topological Layout Learning (ToLL) for 3DSG generation pretraining framework. In detail, we design an Anchor-Conditioned Topological Geometry Reasoning. It adopts a recurrent GNN to recover the global layout of zero-centered subgraphs (the non-visible spatial features) by one anchor with sparse spatial prior. Considering the absence of spatial layout information within the objects, it creates an information bottleneck, compelling our model to recover the full scene layout by leveraging predicate representation learning. Moreover, we construct a Structural Multi-view Augmentation to avoid semantic corruption, enhancing 3DSG representations via self-distillation. The extensive experiments on special dataset demonstrate that our ToLL could often improve 3DSG pertaining quality, outperforming state-of-the-art baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript proposes ToLL, a pretraining framework for 3D Scene Graph (3DSG) generation to mitigate the 'Geometric Shortcut' problem, where models interpolate object positions from dense spatial priors instead of learning topological constraints from edges. The core components are Anchor-Conditioned Topological Geometry Reasoning, which uses a recurrent GNN to recover global layout from zero-centered subgraphs given a single anchor with sparse spatial prior (creating an information bottleneck to force predicate representation learning), and Structural Multi-view Augmentation with asymmetric cross-view self-distillation to preserve semantics. The paper claims this yields improved 3DSG pretraining quality over state-of-the-art baselines on a special dataset.

Significance. If the empirical claims hold with proper validation, the framework could meaningfully advance 3DSG pretraining by enforcing topological reasoning over geometric shortcuts, with potential benefits for generalization in spatial understanding and affordance tasks. The information-bottleneck design and self-distillation approach represent a targeted attempt to address a known limitation in generative pretraining proxies.

major comments (2)
  1. [Abstract] Abstract: The central claim that ToLL 'could often improve 3DSG pretraining quality, outperforming state-of-the-art baselines' is unsupported because the manuscript supplies no quantitative results, error bars, dataset specifications, baseline comparisons, or ablation studies. This absence makes it impossible to evaluate whether the proposed bottleneck actually compels predicate-driven layout recovery rather than residual interpolation.
  2. [Proposed method (Anchor-Conditioned Topological Geometry Reasoning)] Anchor-Conditioned Topological Geometry Reasoning description: The construction of zero-centered subgraphs with a single anchor and sparse spatial prior is asserted to create an information bottleneck that forces use of predicate representations, but no analysis, proof, or ablation is provided to rule out the recurrent GNN exploiting the anchor's residual positional cue plus dataset statistics for interpolation (as noted in the stress-test concern). This assumption is load-bearing for the claim of bypassing geometric shortcuts.
minor comments (3)
  1. [Abstract] The phrase 'special dataset' is imprecise; the manuscript should name the dataset, its scale, and characteristics to allow reproducibility.
  2. [Abstract] Typo: '3DSG pertaining quality' should read '3DSG pretraining quality'.
  3. [Abstract] The term 'often improve' is vague and should be replaced with specific quantitative gains once results are added.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will make revisions to strengthen the presentation of results and analysis.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that ToLL 'could often improve 3DSG pretraining quality, outperforming state-of-the-art baselines' is unsupported because the manuscript supplies no quantitative results, error bars, dataset specifications, baseline comparisons, or ablation studies. This absence makes it impossible to evaluate whether the proposed bottleneck actually compels predicate-driven layout recovery rather than residual interpolation.

    Authors: We agree that the abstract would be strengthened by including key quantitative results. The full manuscript reports extensive experiments with specific performance metrics, error bars from multiple runs, dataset details, baseline comparisons, and ablations. We will revise the abstract to summarize these findings, including the observed improvements over state-of-the-art methods and the dataset used, to better support the claims regarding the information bottleneck. revision: yes

  2. Referee: [Proposed method (Anchor-Conditioned Topological Geometry Reasoning)] Anchor-Conditioned Topological Geometry Reasoning description: The construction of zero-centered subgraphs with a single anchor and sparse spatial prior is asserted to create an information bottleneck that forces use of predicate representations, but no analysis, proof, or ablation is provided to rule out the recurrent GNN exploiting the anchor's residual positional cue plus dataset statistics for interpolation (as noted in the stress-test concern). This assumption is load-bearing for the claim of bypassing geometric shortcuts.

    Authors: We acknowledge that while the manuscript motivates the bottleneck via the zero-centering and sparse prior, additional targeted analysis would help rule out residual geometric exploitation. We will add ablations and stress tests in the revised version, such as randomizing or removing the anchor's positional cues and measuring resulting performance degradation, to empirically demonstrate reliance on predicate representations rather than interpolation from dataset statistics. revision: yes

Circularity Check

0 steps flagged

No circularity: framework introduces independent architectural components without reducing claims to fitted inputs or self-citations

full rationale

The paper's central mechanism (Anchor-Conditioned Topological Geometry Reasoning via recurrent GNN on zero-centered subgraphs plus one sparse anchor) is presented as a new architectural choice that creates an information bottleneck by design. No equations, fitted parameters, or predictions are shown that reduce by construction to prior quantities. The Structural Multi-view Augmentation is likewise introduced as a novel self-distillation step. No self-citation chains, uniqueness theorems, or ansatzes imported from prior author work are invoked as load-bearing justifications. The derivation therefore remains self-contained against external benchmarks and does not collapse to renaming or re-fitting of its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only view supplies no explicit free parameters, axioms, or invented entities; the approach implicitly assumes standard GNN message-passing can recover global layout from sparse anchors and that self-distillation preserves semantics without corruption.

pith-pipeline@v0.9.0 · 5570 in / 1199 out tokens · 33706 ms · 2026-05-14T21:56:03.868043+00:00 · methodology

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 · 41 canonical work pages

  1. [1]

    Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments,

    P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. S ¨underhauf, I. Reid, S. Gould, and A. Van Den Hengel, “Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3674–3683

    work page 2018

  2. [2]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    work page 2023

  3. [3]

    Learning 3d semantic scene graphs from 3d indoor reconstructions,

    J. Wald, H. Dhamo, N. Navab, and F. Tombari, “Learning 3d semantic scene graphs from 3d indoor reconstructions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 3961–3970

    work page 2020

  4. [4]

    3d scene graph: A structure for unified semantics, 3d space, and camera,

    I. Armeni, Z.-Y . He, J. Gwak, A. R. Zamir, M. Fischer, J. Malik, and S. Savarese, “3d scene graph: A structure for unified semantics, 3d space, and camera,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5664–5673

    work page 2019

  5. [5]

    Exploiting edge-oriented reasoning for 3d point-based scene graph analysis,

    C. Zhang, J. Yu, Y . Song, and W. Cai, “Exploiting edge-oriented reasoning for 3d point-based scene graph analysis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9705–9715

    work page 2021

  6. [6]

    Knowledge-inspired 3d scene graph prediction in point cloud,

    S. Zhang, A. Hao, H. Qinet al., “Knowledge-inspired 3d scene graph prediction in point cloud,”Advances in Neural Information Processing Systems, vol. 34, pp. 18 620–18 632, 2021

    work page 2021

  7. [7]

    Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud,

    Z. Wang, B. Cheng, L. Zhao, D. Xu, Y . Tang, and L. Sheng, “Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 21 560–21 569

    work page 2023

  8. [8]

    Clip-driven open- vocabulary 3d scene graph generation via cross-modality contrastive learning,

    L. Chen, X. Wang, J. Lu, S. Lin, C. Wang, and G. He, “Clip-driven open- vocabulary 3d scene graph generation via cross-modality contrastive learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27 863–27 873

    work page 2024

  9. [9]

    Lang3dsg: Language-based contrastive pre-training for 3d scene graph prediction,

    S. Koch, P. Hermosilla, N. Vaskevicius, M. Colosi, and T. Ropinski, “Lang3dsg: Language-based contrastive pre-training for 3d scene graph prediction,” in2024 International Conference on 3D Vision (3DV). IEEE, 2024, pp. 1037–1047

    work page 2024

  10. [10]

    Incremental 3d semantic scene graph prediction from rgb sequences,

    S.-C. Wu, K. Tateno, N. Navab, and F. Tombari, “Incremental 3d semantic scene graph prediction from rgb sequences,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 5064–5074

    work page 2023

  11. [11]

    Hy- perrectangle embedding for debiased 3d scene graph prediction from rgb sequences,

    M. Feng, C. Yan, Z. Wu, W. Dong, Y . Wang, and A. Mian, “Hy- perrectangle embedding for debiased 3d scene graph prediction from rgb sequences,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  12. [12]

    Pointcon- trast: Unsupervised pre-training for 3d point cloud understanding,

    S. Xie, J. Gu, D. Guo, C. R. Qi, L. Guibas, and O. Litany, “Pointcon- trast: Unsupervised pre-training for 3d point cloud understanding,” in European conference on computer vision. Springer, 2020, pp. 574– 591

    work page 2020

  13. [13]

    Sonata: Self-supervised learning of reliable point representations,

    X. Wu, D. DeTone, D. Frost, T. Shen, C. Xie, N. Yang, J. Engel, R. Newcombe, H. Zhao, and J. Straub, “Sonata: Self-supervised learning of reliable point representations,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 193–22 204

    work page 2025

  14. [14]

    Self-supervised pretrain- ing of 3d features on any point-cloud,

    Z. Zhang, R. Girdhar, A. Joulin, and I. Misra, “Self-supervised pretrain- ing of 3d features on any point-cloud,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 252–10 263

    work page 2021

  15. [15]

    arXiv preprint arXiv:2505.12477 , year=

    H. Van Assel, M. Ibrahim, T. Biancalani, A. Regev, and R. Balestriero, “Joint embedding vs reconstruction: Provable benefits of latent space prediction for self supervised learning,”arXiv preprint arXiv:2505.12477, 2025

    work page arXiv 2025

  16. [16]

    Multi-view invariance learning for 3d scene graph pre-training via collaborative cross-modal regulariza- tion,

    Y . Huang, L. Ji, R. Xiao, and J. Sun, “Multi-view invariance learning for 3d scene graph pre-training via collaborative cross-modal regulariza- tion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 7, 2026, pp. 5203–5211

    work page 2026

  17. [17]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in Int. Conf. on Learning Representations, 2014

    work page 2014

  18. [18]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representation learning,”Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  19. [19]

    Sgrec3d: Self-supervised 3d scene graph learning via object-level scene reconstruction,

    S. Koch, P. Hermosilla, N. Vaskevicius, M. Colosi, and T. Ropinski, “Sgrec3d: Self-supervised 3d scene graph learning via object-level scene reconstruction,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 3404–3414

    work page 2024

  20. [20]

    Masked autoencoders for 3d point cloud self-supervised learning,

    Y . Pang, E. H. F. Tay, L. Yuan, and Z. Chen, “Masked autoencoders for 3d point cloud self-supervised learning,”World Scientific Annual Review of Artificial Intelligence, vol. 1, p. 2440001, 2023

    work page 2023

  21. [21]

    A survey on information bottle- neck,

    S. Hu, Z. Lou, X. Yan, and Y . Ye, “A survey on information bottle- neck,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5325–5344, 2024

    work page 2024

  22. [22]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    work page 2021

  23. [23]

    Simplifying dino via coding rate regularization,

    Z. Wu, J. Zhang, D. Pai, X. Wang, C. Singh, J. Yang, J. Gao, and Y . Ma, “Simplifying dino via coding rate regularization,” inForty-second International Conference on Machine Learning, 2025

    work page 2025

  24. [24]

    Graph con- trastive learning with augmentations,

    Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen, “Graph con- trastive learning with augmentations,”Advances in neural information processing systems, vol. 33, pp. 5812–5823, 2020

    work page 2020

  25. [25]

    Graph contrastive learning with adaptive augmentation,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” inProceedings of the web con- ference 2021, 2021, pp. 2069–2080

    work page 2021

  26. [26]

    Graphmae: Self-supervised masked graph autoencoders,

    Z. Hou, X. Liu, Y . Cen, Y . Dong, H. Yang, C. Wang, and J. Tang, “Graphmae: Self-supervised masked graph autoencoders,” inProceed- ings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 594–604

    work page 2022

  27. [27]

    Im- age bert pre-training with online tokenizer,

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong, “Im- age bert pre-training with online tokenizer,” inInternational Conference on Learning Representations

    work page

  28. [28]

    Unsupervised learning of visual features by contrasting cluster assign- ments,

    M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin, “Unsupervised learning of visual features by contrasting cluster assign- ments,”Advances in neural information processing systems, vol. 33, pp. 9912–9924, 2020

    work page 2020

  29. [29]

    Heterogeneous graph learning for scene graph prediction in 3d point clouds,

    Y . Ma, H. Liu, Y . Pei, and Y . Guo, “Heterogeneous graph learning for scene graph prediction in 3d point clouds,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 274–291

    work page 2024

  30. [30]

    3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud,

    M. Feng, H. Hou, L. Zhang, Z. Wu, Y . Guo, and A. Mian, “3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9182–9191

    work page 2023

  31. [31]

    Weakly-supervised 3d scene graph generation via visual-linguistic assisted pseudo-labeling,

    X. Wang, Y . Li, Q. Zhang, W. Wu, M. J. Li, L. Ma, and J. Jiang, “Weakly-supervised 3d scene graph generation via visual-linguistic assisted pseudo-labeling,”IEEE Transactions on Multimedia, vol. 26, pp. 11 164–11 175, 2024

    work page 2024

  32. [32]

    3d scene graph generation from point clouds,

    W. Wei, P. Wei, J. Qin, Z. Liao, S. Wang, X. Cheng, M. Liu, and N. Zheng, “3d scene graph generation from point clouds,”IEEE Trans- actions on Multimedia, vol. 26, pp. 5358–5368, 2023

    work page 2023

  33. [33]

    Concerto: Joint 2d-3d self-supervised learning emerges spatial repre- sentations,

    Y . Zhang, X. Wu, Y . Lao, C. Wang, Z. Tian, N. Wang, and H. Zhao, “Concerto: Joint 2d-3d self-supervised learning emerges spatial repre- sentations,” inThe Thirty-ninth Annual Conference on Neural Informa- tion Processing Systems

    work page

  34. [34]

    Graph self-supervised learning: A survey,

    Y . Liu, M. Jin, S. Pan, C. Zhou, Y . Zheng, F. Xia, and P. S. Yu, “Graph self-supervised learning: A survey,”IEEE transactions on knowledge and data engineering, vol. 35, no. 6, pp. 5879–5900, 2022

    work page 2022

  35. [35]

    Self-supervised learning on graphs: Contrastive, generative, or predictive,

    L. Wu, H. Lin, C. Tan, Z. Gao, and S. Z. Li, “Self-supervised learning on graphs: Contrastive, generative, or predictive,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 4216–4235, 2021

    work page 2021

  36. [36]

    Graph contrastive learning automated,

    Y . You, T. Chen, Y . Shen, and Z. Wang, “Graph contrastive learning automated,” inInternational conference on machine learning. PMLR, 2021, pp. 12 121–12 132

    work page 2021

  37. [37]

    Infogcl: Information-aware graph contrastive learning,

    D. Xu, W. Cheng, D. Luo, H. Chen, and X. Zhang, “Infogcl: Information-aware graph contrastive learning,”Advances in Neural Information Processing Systems, vol. 34, pp. 30 414–30 425, 2021

    work page 2021

  38. [38]

    Het- erogeneous graph masked autoencoders,

    Y . Tian, K. Dong, C. Zhang, C. Zhang, and N. V . Chawla, “Het- erogeneous graph masked autoencoders,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 8, 2023, pp. 9997– 10 005

    work page 2023

  39. [39]

    W. L. Hamilton,Graph representation learning. Morgan & Claypool Publishers, 2020

    work page 2020

  40. [40]

    Point cloud pre-training with diffusion models,

    X. Zheng, X. Huang, G. Mei, Y . Hou, Z. Lyu, B. Dai, W. Ouyang, and Y . Gong, “Point cloud pre-training with diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 935–22 945

    work page 2024

  41. [41]

    Object-centric representation learning for enhanced 3d semantic scene graph prediction,

    K. Heo, G. Kim, S. Kim, and M. Cho, “Object-centric representation learning for enhanced 3d semantic scene graph prediction,” inThe Thirty-ninth Annual Conference on Neural Information Processing Sys- tems, 2025

    work page 2025