Pith. sign in

REVIEW 4 major objections 6 minor 34 references

ODeform claims that 4D object deformation can be modeled as continuous latent flows via two parallel neural ODEs — one for rigid motion, one for local shape change — making arbitrary-time prediction, interpolation, and extrapolation practic

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 ↗

ODeform combines two parallel neural ODEs, one for rigid motion and one for local deformation, to predict arbitrary-time 3D point-cloud deformation from an initial state and physical parameters, outperforming simpler baselines on synthetic benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A reasonable nODE-based deformation architecture, but all quantitative claims rest on self-generated Blender data, and the real-object transfer test leans on synthetic ground truth; worth refereeing, not yet convincing outside the simulator. the 4 major comments →

arxiv 2607.20670 v1 pith:QYJ6YA7Y submitted 2026-07-22 cs.CV

ODeform: Learning Continuous 4D Motion for Shape Deformation with Neural ODEs

Yordanka Velikova , Mahdi Saleh , Liming Kuang , Benjamin Busam This is my paper
classification cs.CV
keywords neural ordinary differential equations4D shape deformationcontinuous dynamicspoint cloud predictionrigid motion decompositionphysical parameter generalizationtemporal interpolationtemporal extrapolation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

ODeform sets out to prove that continuous-time modeling is the right way to predict deformable-object motion, and that it can be done efficiently. The method encodes an initial 3D point cloud plus physical parameters (mass, bending, contact force) into a latent space, then evolves two separate neural ordinary differential equations — one for global rigid motion, one for local deformation — so the object can be reconstructed at any requested timestamp. The paper reports that this dual-ODE design sharply outperforms discrete-step baselines on unseen physical parameters (e.g., RMSE 1.279 mm vs 10.344 mm for a plain neural ODE on unseen mass/bending) and on temporal extrapolation (3.597 mm vs 27.076 mm). If the approach holds up, a single trained model could replace per-scene simulators for robotic manipulation and animation, generating smooth, physically consistent deformations at arbitrary frame rates.

Core claim

The central claim is that the composition of rigid motion and local deformation is the obstacle keeping neural ODEs from working directly on point clouds, and that splitting the dynamics into two parallel latent ODEs removes that obstacle. Since the derivative of a composed transformation T(t)∘D(t)(P0) cannot be expressed as a single vector field over the transformed cloud, the authors evolve T and D separately in latent space, preserving SE(3) structure in the global branch and capturing high-dimensional local shape change in the other. The decoders then recombine the two flows at any time t. The paper argues this yields continuous, physically consistent intermediate states, accurate extrap

What carries the argument

Two parallel neural ODEs initialized by a dual encoder. A compact 32-dimensional latent ODE drives global rigid motion; a 128-dimensional latent ODE drives local deformation. The initial conditions are produced by MLP encoders that take the initial point cloud and physical parameters as input; integration is done with an adaptive Runge-Kutta 4 solver, and MLP decoders map the integrated latents back to an SE(3) pose and point-wise displacements. The key identity is the motion decomposition Φ(P0,θ,t)=T(t)∘D(t)(P0), which motivates the split and keeps the two learned vector fields numerically and semantically separate.

Load-bearing premise

All reported accuracy numbers are measured against ground truth produced by the authors' own physics-engine simulations — contact interactions generated in a soft-body simulator and simulated drop sequences — so the claim that ODeform predicts physically accurate deformation rests on those simulations faithfully representing real soft-object mechanics.

What would settle it

Take real 3D captures of a soft object with known mass and bending stiffness (e.g., a tracked dropped ball or a squeezed pillow with independent material measurements). If ODeform's prediction error against real scans is no better than a discrete-step baseline trained on the same real data — or is dominated by simulator-to-real gap — then the central claim of continuous, physically consistent generalization is not supported.

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

If this is right

  • Prediction at arbitrary timestamps from a single initial frame and physical parameters, without retraining on intermediate observations.
  • Stronger temporal extrapolation than single-ODE or recurrent baselines: on the mass-elastic dataset ODeform reports RMSE 3.597 mm versus 27.076 mm for plain neural ODE when trained on frames 0–20 and tested on the next 10.
  • Robust generalization to unseen physical parameter combinations: 1.279 mm RMSE on unseen mass/bending versus 10.344 mm for the plain neural ODE baseline.
  • Direct transfer to other 3D representations (meshes, gaussian splats) and to real captured objects of the same category, enabling rendering-based animation from one model.
  • Inverse use: recovering material parameters from observed deformation via gradient descent on frozen network weights, with reported MAE 0.113 for mass and 0.084 for bending.

Where Pith is reading between the lines

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

  • The dual-ODE split suggests a modular training scheme the paper does not test: the rigid-motion branch may transfer across object categories while only the deformation branch adapts, which would cut data needs for new objects.
  • Because the ODE is deterministic from initial state, the current formulation cannot represent stochastic or contact-dependent branchings (e.g., an object bouncing one way or another); modeling uncertainty would require a probabilistic ODE or conditioning on contact events, an extension the paper leaves implicit.
  • The successful gaussian-splat decoding points toward real-time render-and-predict loops for robot simulators, but the paper reports no wall-clock timing; measuring runtime against FEM baselines would make the computational-efficiency claim directly testable.
  • Reported error metrics are computed against simulator-generated ground truth; a real-capture benchmark with independently measured material parameters would be the decisive test of whether the learned latent dynamics reflect actual soft-body mechanics.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. ODeform proposes a latent-space neural-ODE formulation for continuous 4D deformation of 3D point clouds. A dual encoder maps an initial point cloud and physical parameters to global and local latent codes; two parallel neural ODEs evolve these codes continuously in time; decoders reconstruct SE(3) pose and point-wise displacements. The method is evaluated on two self-generated Blender datasets (contact-force and mass-elastic), for unseen physical parameters, extrapolation, sparse-frame interpolation, transfer to meshes from HouseCAT6D, integration with 3D Gaussian Splatting, and inverse parameter recovery. The paper claims consistent improvements over nODE, RNN, PEGNN, and OccuFlow baselines, with the largest gains on mass-bending generalization and extrapolation.

Significance. The idea of decomposing rigid and local deformation into two neural ODEs in a learned latent space is physically reasonable and practically attractive for real-time prediction at arbitrary time points. If the reported accuracy held on physically meaningful data, the method would be a useful contribution to deformation prediction for manipulation and simulation. The interpolation from sparse supervision and the inverse parameter-identification experiments are valuable additions. However, the quantitative core is entirely based on the authors' own Blender simulations, and the "real-world transfer" test uses synthetic deformation as ground truth. The physical-accuracy and generalization claims therefore remain unverified with respect to real deformable-object mechanics. No code, data, or uncertainty estimates are provided, limiting reproducibility and comparison.

major comments (4)
  1. [IV-A, Tables I-IV] All quantitative results are evaluated against ground truth generated by the authors' own Blender simulations: the Contact-Force dataset is "enhanced ... using Blender's physics engine" and the Mass-Elastic dataset is generated by simulated drops. The experiment on "real 3D captured objects" (Section IV-D) compares predictions to "ground truth deformation from the synthetic example", not to measured real deformation. Thus the reported RMSE/MAE values measure fidelity to one simulator, not to real soft-object mechanics. For the central "physically accurate, generalizable" claim, at least one validation with real measured deformation fields or a well-established independent simulator is needed.
  2. [III-A, Eq. (3)] The inequality in Eq. (3) is stated without proof and is not mathematically justified as written. For any deterministic differentiable trajectory h(t)=T(t)∘D(t)(P0), the derivative dh/dt can be expressed as f(h(t),t) with f(h,t)=dh/dt on the trajectory. The difficulty is not an inherent impossibility for a single ODE vector field, but a learning/representation challenge. Please reformulate this as a practical motivating observation and support it with an ablation comparing a single ODE against the proposed dual-ODE architecture with the same encoders/decoders.
  3. [Tables I-IV] No uncertainty estimates, number of random seeds, or test-set sizes are reported. Without standard deviations over seeds and explicit ranges of the held-out physical parameters, it is difficult to assess whether the large reported gaps (e.g., RMSE 1.279 vs 10.344 on unseen mass-bending) are stable and meaningful. Report means and stds over at least 3-5 seeds, and state the exact parameter ranges and sampling procedures for the "unseen" conditions.
  4. [IV-C, Table I] The baseline comparison is incomplete for the stated claim. Recent continuous-time methods discussed in the related work (CaSPR, EulerFlow, Modulated nODEs) are not compared, and PEGNN [33] is evaluated only on the Contact-Force dataset, not on Mass-Elastic. In addition, no ablation isolates the contribution of the dual-ODE decomposition, the latent encoders, or the adaptive loss weighting. Adding such ablations and at least one recent continuous baseline would make the empirical claims more convincing.
minor comments (6)
  1. [IV-D] The caption and text for Figure 5 are vague: specify how the model is applied to HouseCAT6D meshes (point sampling, normalization, which physical parameters are used) and what "ground truth deformation from the synthetic example" means exactly.
  2. [IV-G, Table V] Report the parameter ranges and units for mass and bending. An absolute MAE of 0.113 for mass and 0.084 for bending is not interpretable without knowing whether these are normalized values or absolute kg/(N·m²) units.
  3. [III-F] Notation is inconsistent: D is used for the decoder in Eq. (8) and for the deformation mapping D(t) in Eq. (4). Use different symbols (e.g., Dec and Def) to avoid ambiguity.
  4. [III-E] "adaptive Runge-Kutta 4 solver" is unusual: RK4 is a fixed-step method. Specify the actual solver (e.g., dopri5) and tolerances, or remove "adaptive" if a fixed-step RK4 is used.
  5. [References [11],[33]] PE-GNN appears both as related work and as a baseline without noting that it shares a co-author with the current manuscript. A disclosure in the paper would be appropriate.
  6. [II-C] Minor grammatical issues ("Other works has expanded", "present significant theoretical challenges") should be corrected.

Circularity Check

0 steps flagged

No circular derivation: held-out generalization tests drive the claims; only a minor, non-load-bearing self-citation and synthetic-only validation caveats are present.

full rationale

The paper's derivation chain is an empirical encoder-ODE-decoder pipeline, not a mathematical derivation whose output is built into its inputs. The reported results come from held-out splits: unseen physical parameters (Sec. IV-C), extrapolation to later frames (Sec. IV-F, Table III), and sparse-frame interpolation (Table IV). In each case the test data are not used to fit the model, so the low errors are evidence of generalization rather than tautology. The dual-ODE decomposition (Eqs. 6-7) is an architectural choice; no equation maps a fitted parameter directly onto the headline metric. The claim in Eq. (3) that a single nODE cannot model the composed dynamics is mathematically questionable, but it is an architectural motivation, not a circular reduction. PE-GNN [33] is a self-citation by co-author M. Saleh and is used as a baseline/dataset, but it is not load-bearing: the central claims also stand against the non-self-cited nODE and RNN baselines, and the dataset is extended with new Blender simulations rather than used as a proof of correctness. The use of Blender/drop simulations as ground truth and the synthetic ground-truth comparison for real captured meshes (Sec. IV-D) are external-validity limitations, not circularity: the model could have failed these tests, and the synthetic ground truth is not recovered from the model's own output by construction. Accordingly, there is no significant circularity; the score reflects only the minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The learned model relies on standard ODE machinery and on the assumption that deformations are deterministic functions of initial state plus physical parameters. All quantitative claims also assume Blender's rigid-soft body simulation is a trustworthy ground truth. The most fragile premise is Eq. (3), which asserts without proof that a single ODE cannot represent the composed rigid-plus-local motion; this is used to justify the dual-ODE architecture.

free parameters (4)
  • zg latent dimension (32) and zl latent dimension (128) = 32, 128
    Chosen by hand for global vs local encoders; no ablation justifies the values.
  • Encoder depth (5 MLP layers) and decoder depth (3 MLP layers) = 5 and 3
    Architecture choices not justified by experiments.
  • Adaptive loss weights λ1 and λ2 = learned per timestep
    Multi-task weighting from [32]; affects balance between rigid and point losses.
  • ODE solver tolerance / RK4 step choices = not specified
    No details given on numerical solver tolerance; affects accuracy and speed.
axioms (6)
  • domain assumption Deformation dynamics are deterministic given initial state and physical parameters
    Assumed in the Introduction: nODEs are suited because 'each configuration has a unique outcome'; used to justify the ODE formulation.
  • domain assumption Blender's rigid-soft body simulation provides physically correct ground truth
    Section IV-A: all training/test labels for the contact-force and mass-elastic datasets come from Blender; no validation against real measurements.
  • domain assumption The composed deformation can be split as T(t)∘D(t) and separately encoded/decoded
    Section III-C assumes the mapping Φ decomposes into a global SE(3) transform and local deformation.
  • ad hoc to paper Eq. (3) inequality: a single ODE cannot capture the composed derivative
    Stated without proof in Section III-A; motivates the dual-ODE architecture but is not a formal result.
  • domain assumption Point cloud correspondence across time is available
    The decoder predicts point-wise displacements D(t), requiring matching points between frames; this is not discussed.
  • domain assumption The learned latent vector fields can represent the true deformation flow for unseen parameters and geometries
    Universality of the MLP/ODE model is presumed; no generalization bound is offered.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ODeform: Learning Continuous 4D Motion for Shape Deformation with Neural ODEs." pith.science (2026). https://pith.science/paper/QYJ6YA7Y

@misc{pith2026260720670,
  author       = {Pith},
  title        = {Pith review of: ODeform: Learning Continuous 4D Motion for Shape Deformation with Neural ODEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYJ6YA7Y}},
  note         = {Machine review of arXiv:2607.20670}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Modeling continuous object deformation is important for many computer vision and robotics tasks, such as manipulation and simulation. Existing approaches rely on learning-based methods or physics simulators to model shape deformations. However, these approaches either use discrete time steps or are too computationally intensive for real-time applications. We present ODeform, a novel extension of Neural Ordinary Differential Equations to continuous 4D dynamics of deformable objects in 3D space. Our method transforms 3D point clouds and physical conditions (like material properties) into a unified latent space. By solving the resulting ordinary differential equations over time, we model deformations as continuous flows within this learned embedding, eliminating the need for discrete time steps while maintaining computational efficiency. We evaluate our approach on unseen physical parameter configurations, showing improved motion prediction accuracy over baseline methods. Our experiments further demonstrate a successful transfer to real 3D captured objects with novel shapes, along with effective interpolation and extrapolation of the learned dynamics. Our code and data will be made publicly available.

Figures

Figures reproduced from arXiv: 2607.20670 by Benjamin Busam, Liming Kuang, Mahdi Saleh, Yordanka Velikova.

Figure 1
Figure 1. Figure 1: Overview of our continuous deformation modeling framework. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ODeform architecture for continuous deformation modeling. The framework consists of three main components: (1) A dual encoder that [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of mass and bending on object deformation. Inference [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Deformation to real-world captured meshes. The comparison [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 3D Gaussian Splatting integration with ODeform. The visualizations [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Extrapolation results on the FlipFlop object. Trained on first 10 [PITH_FULL_IMAGE:figures/full_fig_p007_7.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

34 extracted references · 5 linked inside Pith

  1. [1]

    Subspace dynamic simulation using rotation-strain coordinates,

    Z. Pan, H. Bao, and J. Huang, “Subspace dynamic simulation using rotation-strain coordinates,”ACM Transactions on Graphics (TOG), vol. 34, pp. 1 – 12, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:16624447

    2015

  2. [2]

    Real-time subspace integration for st. venant-kirchhoff deformable models,

    J. Barbi ˇc and D. L. James, “Real-time subspace integration for st. venant-kirchhoff deformable models,” ACM SIGGRAPH 2005 Papers, 2005. [Online]. Available: https://api.semanticscholar.org/CorpusID:3430137

    2005

  3. [3]

    Pharr, W

    M. Pharr, W. Jakob, and G. Humphreys,Physically based rendering: From theory to implementation. MIT Press, 2023

    2023

  4. [4]

    Meteornet: Deep learning on dynamic 3d point cloud sequences,

    X. Liu, M. Yan, and J. Bohg, “Meteornet: Deep learning on dynamic 3d point cloud sequences,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2019, pp. 9246–9255

    2019

  5. [5]

    Rfnet-4d: Joint object reconstruction and flow estimation from 4d point clouds,

    T.-A. Vu, D. T. Nguyen, B.-S. Hua, Q.-H. Pham, and S.-K. Yeung, “Rfnet-4d: Joint object reconstruction and flow estimation from 4d point clouds,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 36–52

    2022

  6. [6]

    Flownet3d: Learning scene flow in 3d point clouds,

    X. Liu, C. R. Qi, and L. J. Guibas, “Flownet3d: Learning scene flow in 3d point clouds,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 529–537

    2019

  7. [7]

    Deformation and correspondence aware unsupervised synthetic-to-real scene flow esti- mation for point clouds,

    Z. Jin, Y . Lei, N. Akhtar, H. Li, and M. Hayat, “Deformation and correspondence aware unsupervised synthetic-to-real scene flow esti- mation for point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7233–7243

    2022

  8. [8]

    Neu- ral ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neu- ral ordinary differential equations,”Advances in neural information processing systems, vol. 31, 2018

    2018

  9. [9]

    Data-free learning of reduced-order kinematics,

    N. Sharp, C. Romero, A. Jacobson, E. V ouga, P. Kry, D. I. Levin, and J. Solomon, “Data-free learning of reduced-order kinematics,” inACM SIGGRAPH 2023 Conference Proceedings, 2023, pp. 1–9

    2023

  10. [10]

    Robot see robot do: Imitating articulated ob- ject manipulation with monocular 4d reconstruction,

    J. Kerr, C. M. Kim, M. Wu, B. Yi, Q. Wang, K. Goldberg, and A. Kanazawa, “Robot see robot do: Imitating articulated ob- ject manipulation with monocular 4d reconstruction,”arXiv preprint arXiv:2409.18121, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Physics- encoded graph neural networks for deformation prediction under contact,

    M. Saleh, M. Sommersperger, N. Navab, and F. Tombari, “Physics- encoded graph neural networks for deformation prediction under contact,”arXiv preprint arXiv:2402.03466, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    3d-physnet: Learning the intuitive physics of non-rigid object defor- mations,

    Z. Wang, S. Rosa, B. Yang, S. Wang, N. Trigoni, and A. Markham, “3d-physnet: Learning the intuitive physics of non-rigid object defor- mations,”arXiv preprint arXiv:1805.00328, 2018

    Pith/arXiv arXiv 2018

  13. [13]

    Modeling of deformable objects for robotic manipulation: A tutorial and review,

    V . E. Arriola-Rios, P. Guler, F. Ficuciello, D. Kragic, B. Siciliano, and J. L. Wyatt, “Modeling of deformable objects for robotic manipulation: A tutorial and review,”Frontiers in Robotics and AI, vol. 7, p. 82, 2020

    2020

  14. [14]

    Learning contact corrections for handle-based subspace dynamics,

    C. Romero, D. Casas, J. P ´erez, and M. Otaduy, “Learning contact corrections for handle-based subspace dynamics,”ACM Transactions on Graphics (TOG), vol. 40, no. 4, pp. 1–12, 2021

    2021

  15. [15]

    Contact-centric deformation learning,

    C. Romero, D. Casas, M. M. Chiaramonte, and M. A. Otaduy, “Contact-centric deformation learning,”ACM Transactions on Graph- ics (TOG), vol. 41, no. 4, pp. 1–11, 2022

    2022

  16. [16]

    Dynamic hyperbolic attention network for fine hand-object reconstruction,

    Z. Leng, S. cheng Wu, M. Saleh, A. Montanaro, H. Yu, Y . Wang, N. Navab, X. Liang, and F. Tombari, “Dynamic hyperbolic attention network for fine hand-object reconstruction,”2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 14 848–14 858, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:261556700

    2023

  17. [17]

    Tacto: A fast, flexible, and open-source simulator for high-resolution vision-based tactile sensors,

    S. Wang, M. Lambeta, P. wei Chou, and R. Calandra, “Tacto: A fast, flexible, and open-source simulator for high-resolution vision-based tactile sensors,”IEEE Robotics and Automation Letters, vol. 7, pp. 3930–3937, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:229181032

    2020

  18. [18]

    Automatic 3-d manipulation of soft objects by robotic arms with an adaptive deformation model,

    D. Navarro-Alarc ´on, H. M. Yip, Z. Wang, Y . Liu, F. Zhong, T. Zhang, and P. Li, “Automatic 3-d manipulation of soft objects by robotic arms with an adaptive deformation model,”IEEE Transactions on Robotics, vol. 32, pp. 429–441, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:11410308

    2016

  19. [19]

    Fem-based deformation control for dexterous manipulation of 3d soft objects,

    F. Ficuciello, A. Migliozzi, E. Coevoet, A. C. Petit, and C. Duriez, “Fem-based deformation control for dexterous manipulation of 3d soft objects,”2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4007–4013, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:57753541

    2018

  20. [20]

    As-rigid-as-possible surface modeling,

    O. Sorkine and M. Alexa, “As-rigid-as-possible surface modeling,” inSymposium on Geometry processing, vol. 4. Citeseer, 2007, pp. 109–116

    2007

  21. [21]

    Occupancy flow: 4d reconstruction by learning particle dynamics,

    M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger, “Occupancy flow: 4d reconstruction by learning particle dynamics,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5379–5389

    2019

  22. [22]

    Geometry pro- cessing with neural fields,

    G. Yang, S. Belongie, B. Hariharan, and V . Koltun, “Geometry pro- cessing with neural fields,”Advances in Neural Information Processing Systems, vol. 34, pp. 22 483–22 497, 2021

    2021

  23. [23]

    Neural flows: Efficient alternative to neural odes,

    M. Bilo ˇs, J. Sommer, S. S. Rangapuram, T. Januschowski, and S. G ¨unnemann, “Neural flows: Efficient alternative to neural odes,” Advances in neural information processing systems, vol. 34, pp. 21 325–21 337, 2021

    2021

  24. [24]

    Caspr: Learning canonical spatiotemporal point cloud representa- tions,

    D. Rempe, T. Birdal, Y . Zhao, Z. Gojcic, S. Sridhar, and L. J. Guibas, “Caspr: Learning canonical spatiotemporal point cloud representa- tions,”Advances in neural information processing systems, vol. 33, pp. 13 688–13 701, 2020

    2020

  25. [25]

    Graph neural ordinary differential equations,

    M. Poli, S. Massaroli, J. Park, A. Yamashita, H. Asama, and J. Park, “Graph neural ordinary differential equations,”arXiv preprint arXiv:1911.07532, 2019

    Pith/arXiv arXiv 1911

  26. [26]

    Hamiltonian neural networks,

    S. Greydanus, M. Dzamba, and J. Yosinski, “Hamiltonian neural networks,”Advances in neural information processing systems, vol. 32, 2019

    2019

  27. [27]

    Modulated neural odes,

    I. A. Auzina, C ¸ . Yıldız, S. Magliacane, M. Bethge, and E. Gavves, “Modulated neural odes,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  28. [28]

    Ode2vae: Deep gener- ative second order odes with bayesian neural networks,

    C. Yildiz, M. Heinonen, and H. Lahdesmaki, “Ode2vae: Deep gener- ative second order odes with bayesian neural networks,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  29. [29]

    Neural Eulerian Scene Flow Fields,

    K. Vedder, N. Peri, I. Khatri, S. Li, E. Eaton, M. Kocamaz, Y . Wang, Z. Yu, D. Ramanan, and J. Pehserl, “Neural Eulerian Scene Flow Fields,”arXiv preprint arXiv:2410.02031, 2024

    arXiv 2024

  30. [30]

    Neural implicit representations for physical parameter inference from a single video,

    F. Hofherr, L. Koestler, F. Bernard, and D. Cremers, “Neural implicit representations for physical parameter inference from a single video,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 2093–2103

    2023

  31. [31]

    Neural ordinary differential equations,

    T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,” inNeural Information Processing Systems, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:49310446

    2018

  32. [32]

    Auxiliary tasks in multi-task learning,

    L. Liebel and M. K ¨orner, “Auxiliary tasks in multi-task learning,”ArXiv, vol. abs/1805.06334, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:21730434

    Pith/arXiv arXiv 2018

  33. [33]

    Physics- encoded graph neural networks for deformation prediction under contact,

    M. Saleh, M. Sommersperger, N. Navab, and F. Tombari, “Physics- encoded graph neural networks for deformation prediction under contact,”2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 17 160–17 166, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:267499722

    2024

  34. [34]

    Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios,

    H. Jung, S.-C. Wu, P. Ruhkamp, G. Zhai, H. Schieber, G. Rizzoli, P. Wang, H. Zhao, L. Garattoni, S. Meieret al., “Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 498–22 508

    2024

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