Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

The paper claims that per-Gaussian translation-rotation dynamics, learned from multi-view video alone, extrapolate future frames of complex dynamic scenes.

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 ↗

TRACE predicts future frames of dynamic 3D scenes by learning a per-particle translation-rotation dynamics system inside 3D Gaussian Splatting, without labels.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Novel per-Gaussian translation-rotation dynamics for 3DGS with strong held-out extrapolation results, but the 'physics' claim is under-validated: no 3D trajectory ground-truth, no error bars, and the method leans heavily on a deformation-field teacher at training time. the 4 major comments →

arxiv 2508.09811 v1 pith:IERTFLKF submitted 2025-08-13 cs.CV cs.AIcs.CEcs.LGcs.RO

TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos

Jinxi Li , Ziyang Song , Bo Yang This is my paper
classification cs.CV cs.AIcs.CEcs.LGcs.RO
keywords future frame extrapolation3D Gaussian Splattingtranslation rotation dynamicsrigid particledynamic scene reconstructionunsupervised motion segmentationmulti-view videophysics learning
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

The paper aims to show that the motion physics of a dynamic 3D scene can be learned directly from multi-view RGB video, with no object masks, types, or physics equations as supervision. Its proposal is to treat every 3D Gaussian kernel as a rigid particle and learn, for each particle, a complete translation-rotation dynamics system: the rotation center's velocity and acceleration plus the particle's own rotation vector and angular acceleration. Future frames are produced by integrating these learned parameters forward with a second-order Runge-Kutta scheme, while an auxiliary deformation field guides training. If the claim is right, the same parameters that predict the future also cluster into objects and parts by their motion, giving free unsupervised motion segmentation.

Core claim

TRACE's central claim is that a per-particle second-order translation-rotation dynamics system, learned from rendered-image reconstruction losses alone, carries enough physics to extrapolate dynamic scenes beyond the observed time interval. Each canonical 3D Gaussian is treated as a rigid particle with size and orientation; an MLP outputs its rotation-center equivalent velocity, equivalent acceleration, rotation vector, and angular acceleration at a queried time. From those parameters, the RK2 integrator derives future positions and orientations, and the rendered Gaussians are compared with training images. The paper reports that on four datasets the method beats velocity-field and deformati

What carries the argument

The translation rotation dynamics system $\Phi_{\mathrm{Phys}}$: for each 3D Gaussian viewed as a rigid particle, an MLP predicts the equivalent velocity $\tilde{\mathbf{v}}_i$ and acceleration $\tilde{\mathbf{a}}_i$ of its rotation center and the particle's rotation vector $\boldsymbol{\omega}_i$ and angular acceleration $\dot{\boldsymbol{\omega}}_i$. The composite velocity couples the center translation with rotational motion about the center, and a second-order Runge-Kutta integrator using Rodrigues' formula advances the rotation state. The auxiliary deformation field $\Phi_{\mathrm{Def}}$ provides training-time Gaussian trajectories that stabilize early optimization; removing it collapse

Load-bearing premise

The load-bearing premise is that the physical parameters learned with the stabilization of the auxiliary deformation field describe the actual motion law, so they remain valid for times beyond training when the deformation field is clamped and no new observations are available.

What would settle it

Take a scene whose motion changes qualitatively at the end of the training interval, such as an object that moves at constant speed in observed frames and then collides, stops, or reverses. If TRACE's extrapolated frames show the pre-collision motion continuing smoothly, the learned dynamics are an interpolation of observed motion rather than a physical law; the paper reports no such discontinuity test.

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

If this is right

  • Future frames can be extrapolated without PINN losses or object labels; on the new Dynamic Multipart set the reported PSNR is 33.481 versus 28.455 for the strongest velocity-field baseline.
  • Clustering the learned physical parameters yields motion-based object and part segmentation, reaching 95.82 AP on the Dynamic Indoor Scene dataset.
  • The framework is agnostic to the auxiliary deformation backbone: swapping DefGS for 4DGS still extrapolates well, indicating the dynamics module supplies the extrapolation capability.
  • Continual re-training with new observations lets the model track rapidly changing dynamics, with extrapolation PSNR staying around 27-28 as the training window grows.

Where Pith is reading between the lines

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

  • If the learned parameters are treated as state estimates rather than render-only quantities, they could feed differentiable control or trajectory planning for robots, since each particle carries its own acceleration and angular terms.
  • The paper leaves impacts and contacts unmodeled; a natural extension is contact-aware integration or detecting discontinuities from prediction errors, since the RK2 scheme assumes smooth second-order motion.
  • The strong dependence on the auxiliary deformation field (removing it drops Dynamic Multipart PSNR from 33.481 to 19.206) suggests the physics module is underconstrained by images alone; a testable next step is pretraining $\Phi_{\mathrm{Phys}}$ on simulated trajectories or adding multi-frame consistency losses.
  • Clustering parameters instead of RGB or flow could directly support label-free editing, such as moving one segment along its learned trajectory, though the paper only demonstrates segmentation.
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. The paper presents TRACE, a framework for modeling dynamic 3D scenes from multi-view RGB video without labels. It represents the scene with canonical 3D Gaussians and learns, for each Gaussian (treated as a rigid particle), a per-particle translation–rotation dynamics system: rotation-center position/velocity/acceleration and particle rotation vector/angular acceleration, predicted by an MLP (ΦPhys). An auxiliary deformation field (ΦDef) provides training-time motion supervision, and RK2 integration (Algorithm 1) propagates the learned parameters to future times. The method is evaluated on four datasets for future-frame extrapolation (PSNR/SSIM/LPIPS), motion segmentation by clustering learned parameters, and continual learning. The central claim is that ΦPhys learns a genuine, complete set of physical parameters that generalize beyond the training horizon, enabling accurate future frame prediction.

Significance. If the central claim were fully validated, TRACE would be a notable advance: it combines the efficiency of 3DGS with explicit per-particle dynamics, reports large extrapolation gains over strong baselines (e.g., 3–9 dB over NVFi and DefGS+NVFi in Table 1), and demonstrates a clean unsupervised segmentation property by clustering learned parameters. The authors also release code and datasets, which is valuable. However, the claimed physical content of the learned parameters is not directly tested; the current evidence is image-space and horizon-limited. The paper is therefore significant for the empirical system and potential, but its strongest conceptual claim needs additional experimental support.

major comments (4)
  1. [§4.1 / §3.2, Eqs. (3)–(5)] The central claim is that ΦPhys learns a genuine per-particle translation–rotation dynamics system with a complete set of physical parameters. The only quantitative support is rendered RGB extrapolation PSNR/SSIM/LPIPS over the last 14 frames. No experiment validates predicted 3D Gaussian trajectories, rotation centers, angular velocities, or accelerations against ground truth, even though the Dynamic Multipart dataset is synthetic and ground-truth dynamics are available. Short-horizon image-space metrics cannot distinguish a kinematic curve fit from true physical extrapolation. Please add trajectory-level evaluation on Dynamic Multipart (e.g., per-Gaussian 3D position error, rotation error, and parameter recovery) and report it at multiple future horizons, including beyond the current 14-frame window.
  2. [§4.4, Table 4 (row 3) and §3.3] Removing the auxiliary deformation field ΦDef collapses extrapolation PSNR on Dynamic Multipart from 33.481 to 19.206. Since ΦDef is clamped at T_train during extrapolation, the future motion is entirely generated by ΦPhys + RK2, which was trained to reproduce ΦDef's short-interval deformations. This makes the method a teacher–student system for the observed horizon. The manuscript should demonstrate that ΦPhys adds extrapolation power beyond a per-Gaussian second-order Taylor fit of ΦDef's training trajectories, and should quantify the horizon at which its advantage persists (see also Appendix §5.14, Fig. 9).
  3. [§4, Tables 1–4] All quantitative results are single runs with no error bars or significance tests. 3DGS optimization and K-means clustering are stochastic, and the headline contribution is the magnitude of extrapolation gains (up to ~9 dB). Reporting variance over multiple seeds and, where appropriate, significance tests is necessary to establish that the improvements are robust rather than incidental to initialization.
  4. [§3.2, Eq. (4) and Appendix §5.3] The equivalent center velocity and acceleration are defined so that the derived composite velocity exactly reproduces the original expression. This is a reparameterization rather than an independent physical constraint. The paper's claim of learning 'a complete set of physical parameters' therefore requires an identifiability argument or an empirical demonstration that the learned parameters correspond to meaningful physical quantities. On the synthetic Dynamic Multipart dataset, the authors could show recovery of known rotation centers/axes or discuss the non-identifiability and its consequences for the segmentation claim.
minor comments (6)
  1. [Eq. (6)] The loss weights λ_def, λ_phys, λ_img are introduced but their values are not specified in the main text; please list them explicitly.
  2. [Algorithm 1] The algorithm contains typesetting errors and broken references (e.g., 'ref to Eq 5', 'Convert quaternion to rotation matrix: � �� ���'). Please rewrite it so it is self-contained and mathematically readable.
  3. [§4.2] The feature vector used for K-means clustering is not readable in the current rendering ('���� ����� ����� ����� ����� ���'). Specify exactly which learned parameters are concatenated.
  4. [Appendix §5.14, Fig. 9] The figure lacks axis labels and units. This is important for the claim about performance decay with prediction horizon; please add them.
  5. [Abstract / Intro] The phrase 'extraordinary performance' is promotional; consider more neutral wording such as 'consistent improvements'.
  6. [§4.4, Ablation (1)] The time-difference ablation reports numbers only for Dynamic Multipart in Table 4; the discussion claims robustness across datasets, so please include the corresponding rows for Dynamic Object and Dynamic Indoor Scene.

Circularity Check

0 steps flagged

No significant circularity: the central future-frame prediction is a genuinely held-out roll-out from parameters fit only to training frames, and the equivalent-velocity reparameterization is a change of variables, not a constructed prediction.

full rationale

TRACE's central claim is learning per-Gaussian translation-rotation dynamics and extrapolating future frames. The evaluation is held-out: training covers timestamps [0, T_train], and extrapolation is measured on the last 14 frames (Section 4, Appendix 5.14), with no future-frame supervision. The auxiliary deformation field ΦDef is trained only on observed frames and is clamped at T_train during extrapolation, so the future prediction is not a re-rendering of a fitted future field. The ablation in Table 4 row (3) shows performance collapses without ΦDef, but that establishes ΦDef as a useful teacher for learning physics parameters, not that the extrapolation reduces to ΦDef's output by construction. The equivalent velocity/acceleration in Eq. (4) and Appendix 5.3 is an explicit reparameterization: the paper states the first-order equivalence 'is naturally obeyed by the definition,' and then proves the second-order identity. This is a coordinate change inside the dynamic model, not a prediction of an external quantity, so it is not circular. The paper's self-citations (NVFi [21], OGC [47], FreeGave [22]) are used for related work, dataset choice, or baseline construction, and none is invoked as a load-bearing uniqueness theorem or as a substitute for the derivation. The motion-segmentation experiments cluster learned parameters and evaluate against ground-truth masks without using those masks during training, so the segmentation claim is also independently evidenced. Appendix 5.14's decay of extrapolation PSNR with time horizon further confirms that the future frames are not trivially determined by the training fit. No step in the derivation reduces, by construction or by self-citation, to the paper's own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on several hand-selected hyperparameters (tau, dynamics order, unreported loss weights, MLP capacities), four domain assumptions about rigidity and 2nd-order kinematics, and one ad hoc training crutch (the deformation field). The invented entities are reparametrizations and latent parameters without independent falsifiable handles. This ledger shows the paper adds a constrained learning scheme rather than a fully derived physical model.

free parameters (4)
  • Training time difference tau = 2 / frame_rate (tau = 2*Delta_t)
    Hand-selected; ablation (Table 4, group 1) shows tau = 2*Delta_t beats Delta_t and 3*Delta_t, so the reported results depend on this choice.
  • Dynamics order (2nd order) = 2
    Modeling choice; ablation (Table 4, group 2) shows 1st and 3rd order give similar PSNR (within about 1 dB), so the 2nd-order assumption is not strongly validated by data, yet it is central to the method.
  • Loss weights in Eq. 6 (lambda_def, lambda_phys, lambda_img) = not reported
    The total loss is displayed as garbled text with no numeric weights; balancing image, deformation, and physics losses is undisclosed hand-tuning that the central training depends on.
  • MLP capacities for Phi_Phys and Phi_Def = 8 layers, 256 hidden units; positional embedding 8 deg for position, 5 deg for time
    Reported in Appendix 5.5 without ablation; these architecture choices define the capacity of the dynamics and deformation modules.
axioms (4)
  • standard math Any rigid particle motion can be decomposed into rotation about a center that itself translates (Chasles' theorem).
    Invoked in Section 3.2 to justify the rotation-center parametrization. True for exact rigid motion, but the paper additionally fixes scale, making each Gaussian a rigid ellipsoid.
  • domain assumption Per-particle 2nd-order kinematics (constant linear and angular acceleration) is sufficient over the extrapolation horizon.
    Section 3.2 introduces center velocity/acceleration and angular acceleration; the paper argues 2nd order captures common motions but provides no quantitative validation of the approximation over longer horizons or complex interactions.
  • domain assumption Each Gaussian kernel remains rigid with unchanged scale during transport.
    Algorithm 1 assigns the same scale to the deformed Gaussian, stated as 'each Gaussian is rigid with the same scale by definition'; opacity and color are also frozen per prior works.
  • ad hoc to paper The auxiliary deformation field provides reliable supervision for learning the dynamics module.
    Section 3.3 and ablation (3) in Table 4 show the method depends on a parallel deformation field; this engineering assumption is introduced specifically to make optimization work and is not derived from physics.
invented entities (2)
  • Equivalent center velocity and acceleration (c_bar, a_bar) no independent evidence
    purpose: Compounded reparametrization of rotation-center translation and rotation-induced displacement to simplify MLP outputs (Eq. 4).
    Defined so that the composite velocity is reproduced by construction (Appendix 5.3); no external falsifiable handle, purely an internal coordinate change.
  • Per-Gaussian rigid particle dynamics (rotation center, rotation vector with angular acceleration) no independent evidence
    purpose: The central learnable dynamics representation for future extrapolation and segmentation.
    The learned parameters are never compared to ground-truth 3D trajectories or physical constants; their existence is evidenced only through rendered PSNR and K-means clusters, so the physics is an inductive bias, not an independently measured entity.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos." pith.science (2026). https://pith.science/paper/IERTFLKF

@misc{pith2026250809811,
  author       = {Pith},
  title        = {Pith review of: TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IERTFLKF}},
  note         = {Machine review of arXiv:2508.09811}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we aim to model 3D scene geometry, appearance, and physical information just from dynamic multi-view videos in the absence of any human labels. By leveraging physics-informed losses as soft constraints or integrating simple physics models into neural nets, existing works often fail to learn complex motion physics, or doing so requires additional labels such as object types or masks. We propose a new framework named TRACE to model the motion physics of complex dynamic 3D scenes. The key novelty of our method is that, by formulating each 3D point as a rigid particle with size and orientation in space, we directly learn a translation rotation dynamics system for each particle, explicitly estimating a complete set of physical parameters to govern the particle's motion over time. Extensive experiments on three existing dynamic datasets and one newly created challenging synthetic datasets demonstrate the extraordinary performance of our method over baselines in the task of future frame extrapolation. A nice property of our framework is that multiple objects or parts can be easily segmented just by clustering the learned physical parameters.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CausalGS: Learning Physical Causality of 3D Dynamic Scenes with Gaussian Representations

    cs.CV 2026-05 unverdicted novelty 5.0

    CausalGS decouples scene kinematics and dynamics from videos via inverse physics inference on Gaussian representations and guides learning with a differentiable simulator to achieve better long-term future frame prediction.

Reference graph

Works this paper leans on

62 extracted references · 55 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    ParticleNeRF: A Particle-Based Encoding for Online Neural Radiance Fields

    Jad Abou-Chakra, Feras Dayoub, and Niko S ¨underhauf. Par- ticlenerf: Particle based encoding for online neural radiance fields. arXiv preprint arXiv:2211.04041, 2022. 7

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  2. [2]

    Fluid Dynamics Network: Topology- Agnostic 4D Reconstruction via Fluid Dynamics Priors

    Daniele Baieri, Stefano Esposito, Filippo Maggioli, and Emanuele Rodol `a. Fluid Dynamics Network: Topology- Agnostic 4D Reconstruction via Fluid Dynamics Priors. arXiv:2303.09871, 2023. 3

    Pith/arXiv arXiv 2023

  3. [3]

    Nerfies: Deformable Neural Radiance Fields

    Jonathan T Barron, Keunhong Park, Steven M Seitz, and Ri- cardo Martin-brualla. Nerfies: Deformable Neural Radiance Fields. ICCV, 2021. 3

    work page 2021

  4. [4]

    Neural Surface Reconstruction of Dynamic Scenes with Monocular RGB-D Camera

    Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural Surface Reconstruction of Dynamic Scenes with Monocular RGB-D Camera. NeurIPS, 2022

    work page 2022

  5. [5]

    HexPlane: A Fast Representa- tion for Dynamic Scenes

    Ang Cao and Justin Johnson. HexPlane: A Fast Representa- tion for Dynamic Scenes. CVPR, 2023. 2, 3

    work page 2023

  6. [6]

    Scaling physics-informed hard constraints with mixture-of- experts

    Nithin Chalapathi, Yiheng Du, and Aditi S Krishnapriyan. Scaling physics-informed hard constraints with mixture-of- experts. ICLR, 2024. 3

    work page 2024

  7. [7]

    Learning Implicit Fields for Generative Shape Modeling

    Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. CVPR, 2019. 2

    2019

  8. [8]

    Schwing, Alexan- der Kirillov, and Rohit Girdhar

    Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. CVPR,

    work page

  9. [9]

    Neu- ral Unsigned Distance Fields for Implicit Function Learning

    Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neu- ral Unsigned Distance Fields for Implicit Function Learning. NeurIPS, 2020. 2

    work page 2020

  10. [10]

    Learn- ing V ortex Dynamics for Fluid Inference and Prediction

    Yitong Deng, Hong-Xing Yu, Jiajun Wu, and Bo Zhu. Learn- ing V ortex Dynamics for Fluid Inference and Prediction. ICLR, 2023. 3

    work page 2023

  11. [11]

    Boot- sTAP: Bootstrapped training for tracking-any-point

    Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Jo˜ao Carreira, and Andrew Zisserman. Boot- sTAP: Bootstrapped training for tracking-any-point. Asian Conference on Computer Vision, 2024. 6

    work page 2024

  12. [12]

    Neural Radiance Flow for 4D View Synthesis and Video Processing

    Yilun Du, Yinan Zhang, and Joshua B Tenenbaum. Neural Radiance Flow for 4D View Synthesis and Video Processing. ICCV, 2021. 3

    work page 2021

  13. [13]

    Fast Dynamic Radiance Fields with Time-Aware Neural V oxels

    Jiemin Fang, Xinggang Wang, and Matthias Nießner. Fast Dynamic Radiance Fields with Time-Aware Neural V oxels. SIGGRAPH Asia, 2022. 2, 3, 5, 6, 7, 8, 9

    work page 2022

  14. [14]

    Learning to Estimate Single-view V olumetric Flow Motions without 3D Supervision

    Erik Franz, Barbara Solenthaler, and Nils Thuerey. Learning to Estimate Single-view V olumetric Flow Motions without 3D Supervision. ICLR, 2023. 3

    work page 2023

  15. [15]

    K-Planes: Explicit Radiance Fields in Space, Time, and Appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-Planes: Explicit Radiance Fields in Space, Time, and Appearance. CVPR,

    work page

  16. [16]

    Dynamic View Synthesis from Dynamic Monocular Video

    Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic View Synthesis from Dynamic Monocular Video. ICCV, 2021. 3

    work page 2021

  17. [17]

    Bi-level Physics-Informed Neural Net- works for PDE Constrained Optimization Using Broyden’s Hypergrandients

    Zhongkai Hao, Chengyang Ying, Hang Su, Jun Zhu, Jian Song, and Ze Cheng. Bi-level Physics-Informed Neural Net- works for PDE Constrained Optimization Using Broyden’s Hypergrandients. ICLR, 2023. 3

    work page 2023

  18. [18]

    Learning to Simu- late Complex Physics with Graph Networks

    Alvaro Sanchez-gonzalez Jonathan, Godwin Tobias, Pfaff Rex, Ying Jure, and Peter W Battaglia. Learning to Simu- late Complex Physics with Graph Networks. ICML, 2020. 1, 3

    work page 2020

  19. [19]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Bernhard Kerbl, Universit ´e C ˆote, Georgios Kopanas, Uni- versit´e Cˆote, Thomas Leimk¨uhler, Max-planck-institut Infor- matik, and G R Aug. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. SIGGRAPH, 2023. 1, 2, 3

    work page 2023

  20. [20]

    MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds

    Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds. arXiv:2405.17421,

    Pith/arXiv arXiv

  21. [21]

    NVFi: Neural Veloc- ity Fields for 3D Physics Learning from Dynamic Videos

    Jinxi Li, Ziyang Song, and Bo Yang. NVFi: Neural Veloc- ity Fields for 3D Physics Learning from Dynamic Videos. NeurIPS, 2023. 2, 5, 6, 7, 3, 8, 9

    work page 2023

  22. [22]

    Free- Gave: 3D Physics Learning from Dynamic Videos by Gaus- sian Velocity

    Jinxi Li, Ziyang Song, Siyuan Zhou, and Bo Yang. Free- Gave: 3D Physics Learning from Dynamic Videos by Gaus- sian Velocity. CVPR, 2025. 2

    work page 2025

  23. [23]

    Neu- ral 3D Video Synthesis From Multi-View Video

    Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, and Zhaoyang Lv. Neu- ral 3D Video Synthesis From Multi-View Video. CVPR,

    work page

  24. [24]

    PAC-NeRF: Physics Augmented Continuum Neural Radiance Fields for Geometry-Agnostic System Identifica- tion

    Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. PAC-NeRF: Physics Augmented Continuum Neural Radiance Fields for Geometry-Agnostic System Identifica- tion. ICLR, 2023. 1

    work page 2023

  25. [25]

    Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes. CVPR, 2021. 3, 8

    work page 2021

  26. [26]

    Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis. CVPR, 2024. 3

    work page 2024

  27. [27]

    From Zero to Turbulence: Generative Modeling for 3D Flow Simulation

    Marten Lienen, David L ¨udke, Jan Hansen-Palmus, and Stephan G¨unnemann. From Zero to Turbulence: Generative Modeling for 3D Flow Simulation. ICLR, 2024. 3

    work page 2024

  28. [28]

    Lawrence Zitnick, and Piotr Doll ´ar

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Doll ´ar. Microsoft COCO: Common Objects in Context. ECCV, 2014. 7

    work page 2014

  29. [29]

    Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaus- sian Particle

    Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaus- sian Particle. CVPR, 2024. 3

    work page 2024

  30. [30]

    Physics3D: Learning Physical Properties of 3D Gaussians via Video Diffusion

    Fangfu Liu, Hanyang Wang, Shunyu Yao, and Shengjun Zhang. Physics3D: Learning Physical Properties of 3D Gaussians via Video Diffusion. arXiv:2406.04338, 2024. 3

    Pith/arXiv arXiv 2024

  31. [31]

    Dynamic Gaus- sians Mesh: Consistent Mesh Reconstruction from Monocu- lar Videos

    Isabella Liu, Hao Su, and Xiaolong Wang. Dynamic Gaus- sians Mesh: Consistent Mesh Reconstruction from Monocu- lar Videos. arXiv:2404.12379, 2024. 3

    Pith/arXiv arXiv 2024

  32. [32]

    Robust Dynamic Radiance Fields

    Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Jo- hannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. CVPR, 2023. 3

    work page 2023

  33. [33]

    3D Geometry-aware Deformable Gaussian Splatting for Dynamic View Synthesis

    Zhicheng Lu, Xiang Guo, Le Hui, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3D Geometry-aware Deformable Gaussian Splatting for Dynamic View Synthesis. CVPR,

    work page

  34. [34]

    Occupancy Net- works: Learning 3D Reconstruction in Function Space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy Net- works: Learning 3D Reconstruction in Function Space. CVPR, 2019. 2

    work page 2019

  35. [35]

    Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 1, 2

    work page 2020

  36. [36]

    Estimates on the generalization error of physics-informed neural networks for approximating PDEs

    Siddhartha Mishra and Roberto Molinaro. Estimates on the generalization error of physics-informed neural networks for approximating PDEs. IMA Journal of Numerical Analysis ,

    work page

  37. [37]

    DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation. CVPR, 2019. 2

    work page 2019

  38. [38]

    Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- Brualla, and Steven M

    Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- Brualla, and Steven M. Seitz. HyperNeRF: A Higher- Dimensional Representation for Topologically Varying Neu- ral Radiance Fields. SIGGRAPH Asia, 2021. 3

    work page 2021

  39. [39]

    Temporal Interpola- tion Is All You Need for Dynamic Neural Radiance Fields

    Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahyup Kang. Temporal Interpola- tion Is All You Need for Dynamic Neural Radiance Fields. CVPR, 2023. 3

    work page 2023

  40. [40]

    D-NeRF: Neural Radiance Fields for Dynamic Scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. CVPR, 2021. 1, 2, 5, 6, 7, 3, 8, 9

    work page 2021

  41. [41]

    Yi-Ling Qiao, Alexander Gao, and Ming C. Lin. Ne- uPhysics: Editable Neural Geometry and Physics from Monocular Videos. NeurIPS, 2022. 3

    work page 2022

  42. [42]

    NeuSmoke: Efficient Smoke Re- construction and View Synthesis with Neural Transportation Fields

    Jiaxiong Qiu, Ruihong Cen, Zhong Li, Han Yan, Ming- ming Cheng, and Bo Ren. NeuSmoke: Efficient Smoke Re- construction and View Synthesis with Neural Transportation Fields. SIGGRAPH Asia, 2024. 1, 3

    work page 2024

  43. [43]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 2019. 1, 2

    work page 2019

  44. [44]

    Hidden fluid mechanics: Learning velocity and pres- sure fields from flow visualizations

    Maziar Raissi, Alireza Yazdani, and George Em Karni- adakis. Hidden fluid mechanics: Learning velocity and pres- sure fields from flow visualizations. Science, 2020. 3

    work page 2020

  45. [45]

    Olinde Rodrigues. Des lois g ´eom´etriques qui r ´egissent les d´eplacements d’un syst `eme solide dans l’espace, et de la variation des coordonn ´ees provenant de ces d ´eplacements consid´er´es ind´ependants des causes qui peuvent les produire. Journal de Math ´ematiques Pures et Appliqu ´ees, 1840. 5

    work page

  46. [46]

    Schonberger and Jan-Michael Frahm

    Johannes L. Schonberger and Jan-Michael Frahm. Structure- from-Motion Revisited. CVPR, 2016. 3

    work page 2016

  47. [47]

    OGC: Unsupervised 3D Ob- ject Segmentation from Rigid Dynamics of Point Clouds

    Ziyang Song and Bo Yang. OGC: Unsupervised 3D Ob- ject Segmentation from Rigid Dynamics of Point Clouds. NeurIPS, 2022. 7, 4

    work page 2022

  48. [48]

    MonoNeRF: Learning a Generalizable Dynamic Radiance Field from Monocular Videos

    Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoNeRF: Learning a Generalizable Dynamic Radiance Field from Monocular Videos. ICCV, 2023. 3

    work page 2023

  49. [49]

    Non- Rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video

    Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollh¨ofer, Christoph Lassner, and Christian Theobalt. Non- Rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video. ICCV, 2021

    work page 2021

  50. [50]

    Neural Trajectory Fields for Dynamic Novel View Synthesis

    Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural Trajectory Fields for Dynamic Novel View Synthesis. arXiv:2105.05994, 2021. 3

    Pith/arXiv arXiv 2021

  51. [51]

    Whitney, Tatiana Lopez-Guevara, Tobias Pfaff, Yulia Rubanova, Thomas Kipf, Kimberly Stachenfeld, and Kelsey R

    William F. Whitney, Tatiana Lopez-Guevara, Tobias Pfaff, Yulia Rubanova, Thomas Kipf, Kimberly Stachenfeld, and Kelsey R. Allen. Learning 3D Particle-based Simulators from RGB-D Videos. ICLR, 2024. 1, 3

    work page 2024

  52. [52]

    4D Gaussian Splatting for Real-Time Dynamic Scene Ren- dering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Ren- dering. CVPR, 2024. 1, 2, 3, 5, 6

    work page 2024

  53. [53]

    Space-time Neural Irradiance Fields for Free- Viewpoint Video

    Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time Neural Irradiance Fields for Free- Viewpoint Video. CVPR, 2021. 3

    work page 2021

  54. [54]

    Tenenbaum, Yunzhu Li, and Hsiao-Yu Tung

    Haotian Xue, Antonio Torralba, Daniel LK Yamins, Joshua B. Tenenbaum, Yunzhu Li, and Hsiao-Yu Tung. 3D- IntPhys: Learning 3D Visual Intuitive Physics for Fluids, Rigid Bodies, and Granular Materials. NeurIPS, 2023. 3

    work page 2023

  55. [55]

    Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 9

    work page 2024

  56. [56]

    Gaussian grouping: Segment and edit anything in 3d scenes

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 6

    work page 2024

  57. [57]

    Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera

    Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera. CVPR, 2020. 5

    work page 2020

  58. [58]

    Decoupling Dynamic Monocular Videos for Dynamic View Synthesis

    Meng You and Junhui Hou. Decoupling Dynamic Monocu- lar Videos for Dynamic View Synthesis. arXiv:2304.01716,

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T

    Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Bran- don Y . Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics- Based Interaction with 3D Objects via Video Generation. arXiv:2404.13026, 2024. 3

    Pith/arXiv arXiv 2024

  60. [60]

    Aditya Prakash

    Zhiyuan Zhao, Xueying Ding, and B. Aditya Prakash. PINNsFormer: A Transformer-Based Framework For Physics-Informed Neural Networks. ICLR, 2024. 3

    work page 2024

  61. [61]

    Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians

    Licheng Zhong, Hong-xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians. ECCV, 2024. 1, 3

    work page 2024

  62. [62]

    Chessboard

    Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EW A volume splatting. IEEE Visualiza- tion, 2001. 1 TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos Supplementary Material The appendix includes: • Rendering equation and preliminary for vanilla 3DGS. • Implementation details of Auxiliary Deformation Field. • ...

    work page arXiv 2001

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.