Pith. sign in

REVIEW 4 major objections 5 minor 51 references

3D Gaussian Representations with Motion Trajectory Field for Dynamic Scene Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Monocular dynamic scenes can be reconstructed for both novel-view synthesis and motion recovery in real time by representing each Gaussian's motion as a global reference point plus a linear combination of shared learnable basis trajectories

desk verdict Rendering results are real and competitive, but the motion trajectory recovery claim is unmeasured and arguably ill-defined under densification; referee it, but demand a trajectory metric and an answer to Gaussian identity. read the letter →

arxiv 2508.07182 v1 pith:KKCYESLM submitted 2025-08-10 cs.RO

Xuesong Li , Lars Petersson , Vivien Rolland This is my paper
classification cs.RO
keywords dynamicscenereconstruction3DGaussianSplattingmotiontrajectoryfieldnovel-viewsynthesismonocularvideolow-rankspacerecoveryregularization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper attempts to solve two problems at once from a single monocular video: synthesize novel views of a dynamic scene and recover the underlying 3D motion. Its proposal is to represent every moving 3D Gaussian not by per-frame positions but by a compact motion trajectory field: each Gaussian keeps a global reference point and a fixed set of coefficients, and its position at any time is the reference point plus a weighted sum of a small number of shared, learnable basis trajectories. Because the bases are shared across all Gaussians, neighboring points move coherently, and because the coefficients are time-invariant, the optimization stays small enough to keep 3D Gaussian Splatting's real-time rendering. If the paper is right, ordinary monocular video can be turned into both renderable 3D scenes and usable motion trajectories, with the best mean rendering metrics reported on the synthetic and real-world benchmarks used.

What carries the argument

The motion trajectory field is the factorization $\mu_t^i = \bar\mu^i + \sum_{k=1}^K c_k^i B_k(t)$, with $K=5$ learnable basis trajectories $B_k(t)$ shared by all Gaussians; per-Gaussian coefficients $c_k^i$ come from a small MLP over the global reference point with positional encoding. This single object carries the argument: it replaces per-frame or per-Gaussian deformation fields with a global low-rank motion prior, enforces coherent neighboring motion through shared bases, keeps the number of optimized parameters small, and lets the renderer stay at 3D Gaussian Splatting speed.

What would settle it

Render a synthetic monocular scene with known per-point ground-truth trajectories (for example, an articulated figure), run the method with K=5, extract every Gaussian's center per frame, and compare against the ground-truth track; if the median positional error is large while novel-view PSNR stays high, the low-rank trajectory field is render-plausible but not physically accurate.

Watch

Extended reading notes

Core claim

At the center of the paper is the claim that dynamic scene reconstruction and motion recovery can be unified through a low-rank motion trajectory field on top of 3D Gaussian Splatting. For each Gaussian $i$, the method writes its position at time $t$ as $\mu_t^i = \bar\mu^i + \sum_{k=1}^K c_k^i B_k(t)$, where $\bar\mu^i$ is a detached global reference point, $B_k(t)$ is one of $K=5$ globally shared basis trajectories over time (initialized as DCT bases and fine-tuned), and $c_k^i$ is a per-Gaussian coefficient produced by a lightweight MLP with positional encoding. The same factorization is extended to time-varying covariance, scale, and rotation bases. The paper argues that this representat

Load-bearing premise

The load-bearing premise is that every dynamic Gaussian's motion lies in a low-rank space spanned by five globally shared, time-invariant basis trajectories; if a scene's true motion cannot be expressed that way, the recovered trajectories can look right in the rendered images but be wrong in the world.

Editorial extensions

If this is right

  • Dynamic scenes from a single monocular video can be rendered at real-time frame rates while matching or exceeding existing baselines on the synthetic and real-world benchmarks.
  • The recovered per-Gaussian trajectories provide a motion field, not just a renderable geometry, so downstream tasks such as robot manipulation and navigation could use them directly.
  • Decoupling static and dynamic Gaussians via segmentation reduces primitive counts and speeds rendering while improving quality, as shown by the ablation listing 740k to 480k Gaussians and 13 to 26 FPS.
  • The DCT-initialized, learnable basis functions give a compact temporal prior, avoiding per-frame optimization and the associated memory and storage costs.

Reading between the lines

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

  • Beyond the paper: the same low-rank trajectory factorization could be evaluated directly on physical motion accuracy by using synthetic scenes with known per-point ground-truth tracks; the paper reports rendering metrics and shows trajectories qualitatively but does not report trajectory error, so that metric is the natural next test.
  • Beyond the paper: the reported ablation shows that adding motion regularization lowers PSNR/SSIM while smoothing the trajectories, so a robotics user should expect a tradeoff between render fidelity and physical plausibility; a trajectory-error metric would tell when the tradeoff is worth it.
  • Beyond the paper: if the five-basis subspace generalizes, scaling to long videos or multi-object scenes could be done by adding more basis functions or per-object basis subsets rather than per-frame optimization; the DCT initialization suggests frequency-ordered bases, making the required rank for a scene a measurable quantity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a dynamic scene reconstruction method based on 3D Gaussian Splatting. Each Gaussian's center is decomposed into a global reference point plus a linear combination of k shared learnable trajectory basis vectors, with per-Gaussian coefficients predicted by an MLP (Eqs. 1--3). The representation also models time-varying covariance with additional bases, separates static/dynamic Gaussian via a learnable state and a pseudo-segmentation mask, and applies ARAP and spatial smoothness regularizers. Experiments on D-NeRF and HyperNeRF report novel-view synthesis metrics (PSNR/SSIM/LPIPS) and qualitative trajectory visualizations. The authors claim state-of-the-art results in both novel-view synthesis and motion trajectory recovery from monocular video.

Significance. If the rendering results are reproducible, the compact shared-basis motion representation is a useful step for real-time dynamic scene rendering with 3DGS, and the static/dynamic decoupling plus motion regularization is a sensible way to constrain an under-determined problem. The paper includes quantitative comparisons on standard benchmarks and an ablation for static/dynamic separation. However, the second headline contribution—physically plausible trajectory recovery—is currently supported only by qualitative images and internal consistency arguments; no trajectory metric is reported. Given the paper's own framing, this gap is substantial.

major comments (4)
  1. [Abstract; §4.2; Figs. 3 and 6] The claim of state-of-the-art 'motion trajectory recovery' is not quantitatively evaluated anywhere. Tables 1 and 2 report only PSNR/SSIM/LPIPS for novel-view synthesis. The only evidence for trajectory recovery is qualitative visualization (Fig. 3, Fig. 6). D-NeRF is synthetic and could provide ground-truth point trajectories or at least a correspondence-based error; comparisons against trajectory-output baselines such as Gaussian-flow or Dynamic 3DGS are also needed. Without a trajectory error metric, the second headline contribution is unsupported.
  2. [§4.1.2; Eqs. (1)–(3)] Adaptive density control splits, clones, and prunes Gaussians until 25k iterations. A trajectory in Eqs. (1)–(3) is defined per Gaussian index i, but after cloning/splitting, index i no longer corresponds to the same physical surface point, and no correspondence preservation mechanism is described. Thus the rendered 'trajectories' may reflect primitive identity rather than physical motion. The paper must either modify/describe the densification to maintain persistent point identity (e.g., by carrying coefficients to offspring) or provide evidence that trajectories are robust to densification, and define what physical quantity is recovered.
  3. [§4.3.2, Table 4] The motion regularization ablation shows that the no-regularization baseline has the highest PSNR (27.4) and SSIM (0.879); all regularized variants are slightly lower. The text argues that regularization yields 'the smoothest trajectory' and 'motion consistency', but no quantitative smoothness or trajectory-accuracy metric is provided. Because these regularizers are a claimed contribution and are central to the 'physically plausible trajectories' claim, the ablation is inconclusive. Report trajectory-oriented metrics (e.g., acceleration error, ARAP residual, correspondence accuracy) in addition to rendering metrics.
  4. [§3.2, §4.1.2] The method fixes k=5 shared trajectory bases (and similarly small numbers for scale/rotation bases) without an ablation or expressiveness study. The low-rank trajectory-space assumption is a strong prior; for complex articulated motions or topology changes, 5 bases may be insufficient, and the rendered trajectories could be plausible yet physically wrong. The paper should include an ablation over k and a synthetic experiment with known non-rigid deformations to test the subspace capacity.
minor comments (5)
  1. [Tables 3 and 4] These tables do not state which dataset/scene the ablation is computed on, nor whether standard deviations over runs are available. The 'Num' row in Table 3 is also ambiguous ('multiplied by 10^3').
  2. [Table 1] Table 1 lacks an average row and error bars; the text says 'outperforms ... in most scenes' but the method loses to D-3DGS on Bouncing Balls and Lego (e.g., PSNR 40.15 vs 41.01 and 30.48 vs 33.07). Please make the claims more precise.
  3. [Eqs. (9) and (10)] The ARAP and spatial smoothness formulas contain undefined or mis-rendered symbols (e.g., the weighting factor and perturbation epsilon). Please rewrite with clear notation and define all variables.
  4. [Throughout] Typographical errors: 'Guassians' (Sec. 3.2), 'cased by' (Sec. 1), 'we firstly' (Sec. 4.1.2), and 'evalution' (Sec. 4.1.1).
  5. [§4.1.1, §4.3.1] The pseudo-segmentation masks are generated with Track Anything [46]; the paper should quantify mask noise or at least discuss failure cases, since segmentation errors propagate to the state parameters and motion field.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the trajectory-recovery claim is under-verified but not derived from its own outputs.

full rationale

The paper's central derivation is an end-to-end optimization of 3D Gaussian parameters (centers expressed as global reference points plus linear combinations of shared DCT-initialized, learnable trajectory bases with MLP coefficients; Eqs. 1-6) against photometric, segmentation, and motion losses. Novel-view synthesis is evaluated on held-out frames of D-NeRF and HyperNeRF, which provides independent external grounding. I found no step where a fitted quantity is renamed as a prediction, no parameter fitted to a subset and then used to predict the same subset, and no load-bearing appeal to the authors' own prior results; the self-citations [20,21,45] are contextual and not load-bearing. The assumption of low-rank trajectory space is imported from external works [1,23,40] and is explicitly stated as an assumption, not derived. The unverified portion is the motion-trajectory recovery claim: it is supported only by qualitative visualizations (Fig. 3, Fig. 6) and by an ablation (Table 4) that actually reports lower PSNR/SSIM with regularization, while no trajectory error metric is reported. Also, since adaptive density control runs until 25k iterations (Sec. 4.1.2), Gaussian identity across splits/clones/prunes is not guaranteed, so 'trajectories' may not track physical points. These are correctness/verification gaps, not circular reductions: the rendering numbers do not depend on the trajectory claim being true. Hence the circularity score is 0.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The central modeling assumption is a low-rank trajectory-space factorization borrowed from NRSfM/NeTF; the paper adds learnable bases, an MLP, masks, and hand-tuned losses. No new physical entity is introduced. The main unverified premise is that photometric supervision plus smoothness priors resolves monocular motion ambiguity, a premise not checked with any quantitative trajectory metric.

free parameters (7)
  • Number of position trajectory bases k = 5
    Hand-chosen in Sec. 3.2; controls how many global motion patterns the field can express.
  • Number of scale and rotation motion bases = not reported
    Eqs. (5)-(6) add separate bases for covariance scale/rotation, but the counts are not specified; they are hyperparameters.
  • Loss weights (seg, state, ARAP, spatial) = 0.3, 0.6, 0.5, 0.5
    Set manually in Sec. 4.1.2; balance rendering quality against segmentation and motion smoothness.
  • Densification gradient thresholds (static, dynamic) = 0.0002, 0.0004
    Hand-set in Sec. 4.1.2; higher dynamic threshold prevents over-densification of moving Gaussians.
  • Static selection mask threshold = 0.8
    Set in Sec. 3.3 to make static/dynamic classification robust to segmentation errors.
  • Spatial smoothness perturbation scale factor = proportional to scene scale, factor not fully specified
    Adaptive perturbation magnitude in Eq. (10) uses a hand-chosen coefficient.
  • Coefficient MLP architecture and positional encoding frequencies = not specified
    The lightweight MLP in Eq. (3) is not described in detail, leaving architecture choices open.
assumptions (7)
  • standard math Trajectory space factorization: time-varying 3D point structure can be factored into basis trajectories and per-point coefficients (Akhter et al., trajectory space).
    Invoked in Sec. 3.1 as the mathematical foundation for the motion trajectory field.
  • domain assumption A small number of basis trajectories (5) is sufficient for the dynamic foreground in these datasets.
    Set in Sec. 3.2; no analysis of basis sufficiency or trajectory reconstruction error is provided.
  • domain assumption DCT-initialized learnable motion bases can represent real-world motions after fine-tuning, adopted from DynIBaR.
    Used in Sec. 3.2 to justify replacing fixed DCT bases with learnable ones.
  • domain assumption Pseudo foreground/background masks from Track Anything are sufficiently accurate for supervising the static/dynamic state of Gaussians.
    Used in Sec. 3.3 to train the per-Gaussian state probability; mask errors are only partially mitigated by a threshold.
  • domain assumption ARAP prior: neighboring Gaussians should move by a rigid transformation between two timesteps.
    Defined in Eq. (9), Sec. 3.4; used to regularize motion of dynamic points.
  • domain assumption Spatial smoothness: perturbing a global reference point should not change the MLP outputs much.
    Defined in Eq. (10), Sec. 3.4; used to enforce smoothness over neighboring spatial locations.
  • ad hoc to paper Photometric loss, mask loss, and smoothness losses are enough to resolve monocular motion ambiguity and recover physically true trajectories.
    This is the paper's key unvalidated premise behind the motion trajectory recovery claim; no quantitative trajectory ground truth is supplied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Gaussian Representations with Motion Trajectory Field for Dynamic Scene Reconstruction." pith.science (2026). https://pith.science/paper/KKCYESLM

@misc{pith2026250807182,
  author       = {Pith},
  title        = {Pith review of: 3D Gaussian Representations with Motion Trajectory Field for Dynamic Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KKCYESLM}},
  note         = {Machine review of arXiv:2508.07182}
}
read the original abstract

This paper addresses the challenge of novel-view synthesis and motion reconstruction of dynamic scenes from monocular video, which is critical for many robotic applications. Although Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have demonstrated remarkable success in rendering static scenes, extending them to reconstruct dynamic scenes remains challenging. In this work, we introduce a novel approach that combines 3DGS with a motion trajectory field, enabling precise handling of complex object motions and achieving physically plausible motion trajectories. By decoupling dynamic objects from static background, our method compactly optimizes the motion trajectory field. The approach incorporates time-invariant motion coefficients and shared motion trajectory bases to capture intricate motion patterns while minimizing optimization complexity. Extensive experiments demonstrate that our approach achieves state-of-the-art results in both novel-view synthesis and motion trajectory recovery from monocular video, advancing the capabilities of dynamic scene reconstruction.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 34 canonical work pages

  1. [1]

    Nonrigid structure from motion in trajectory space.Ad- vances in neural information processing systems, 21, 2008

    Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Nonrigid structure from motion in trajectory space.Ad- vances in neural information processing systems, 21, 2008. 1, 2, 4

    work page 2008

  2. [2]

    Trajectory space: A dual representation for nonrigid struc- ture from motion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442–1456, 2010

    Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid struc- ture from motion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442–1456, 2010. 2

    work page 2010

  3. [3]

    Recovering non-rigid 3d shape from image streams

    Christoph Bregler, Aaron Hertzmann, and Henning Bier- mann. Recovering non-rigid 3d shape from image streams. InProceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690–696. IEEE, 2000. 2

    work page 2000

  4. [4]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130–141, 2023. 1, 2

    2023

  5. [5]

    Neural radiance flow for 4d view synthesis and video processing

    Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenen- baum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 14304–14314. IEEE Computer Society, 2021. 1

    work page 2021

  6. [6]

    Fast dynamic radiance fields with time-aware neural vox- els

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xi- aopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural vox- els. InSIGGRAPH Asia 2022 Conference Papers, pages 1–9,

    work page 2022

  7. [7]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023. 1, 2, 6

    work page 2023

  8. [8]

    V4d: V oxel for 4d novel view synthesis

    Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: V oxel for 4d novel view synthesis. IEEE Transactions on Visualization and Computer Graph- ics, 2023. 2

    work page 2023

Show all 51 references
  1. [9]

    Dynamic view synthesis from dynamic monocular video

    Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712–5721, 2021. 2

    2021

  2. [10]

    Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction.arXiv preprint arXiv:2403.11447, 2024

    Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction.arXiv preprint arXiv:2403.11447, 2024. 2

    2024 arXiv

  3. [11]

    An ef- ficient 3d gaussian representation for monocular/multi-view dynamic scenes.arXiv preprint arXiv:2311.12897, 2023

    Kai Katsumata, Duc Minh V o, and Hideki Nakayama. An ef- ficient 3d gaussian representation for monocular/multi-view dynamic scenes.arXiv preprint arXiv:2311.12897, 2023. 2

    2023 arXiv

  4. [12]

    Segment anything in high qual- ity.Advances in Neural Information Processing Systems, 36,

    Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi- Keung Tang, Fisher Yu, et al. Segment anything in high qual- ity.Advances in Neural Information Processing Systems, 36,

  5. [13]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023. 1, 2, 3, 4, 5, 6

    2023

  6. [14]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations (ICLR), San Diega, CA, USA,

  7. [15]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 4

    2023

  8. [16]

    Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting.arXiv preprint arXiv:2312.00112, 2023

    Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting.arXiv preprint arXiv:2312.00112, 2023. 2

    2023 arXiv

  9. [17]

    Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames

    Suryansh Kumar, Yuchao Dai, and Hongdong Li. Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames. InProceedings of the IEEE inter- national conference on computer vision, pages 4649–4657,

  10. [18]

    Spatio- temporal union of subspaces for multi-body non-rigid structure-from-motion.Pattern Recognition, 71:428–443,

    Suryansh Kumar, Yuchao Dai, and Hongdong Li. Spatio- temporal union of subspaces for multi-body non-rigid structure-from-motion.Pattern Recognition, 71:428–443,

  11. [19]

    Neural 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, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vi- si...

    2022

  12. [20]

    Xuesong Li and Jose E Guivant. Efficient and accurate object detection with simultaneous classification and tracking under limited computing power.IEEE Transactions on Intelligent Transportation Systems, 24(6):5740–5751, 2023. 1

    2023

  13. [21]

    Real-time 3d object proposal generation and classification using limited processing resources.Robotics and Autonomous Systems, 130:103557, 2020

    Xuesong Li, Jose Guivant, and Subhan Khan. Real-time 3d object proposal generation and classification using limited processing resources.Robotics and Autonomous Systems, 130:103557, 2020. 1

    2020

  14. [22]

    Neural scene flow fields for space-time view synthesis of dy- namic scenes

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dy- namic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498– 6508, 2021. 2

    2021

  15. [23]

    Dynibar: Neural dynamic image-based rendering

    Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4

    2023

  16. [24]

    Generative image dynamics

    Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142–24153, 2024. 3

    2024

  17. [25]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136– 21145, 2024. 1, 2, 6

    2024

  18. [26]

    Modgs: Dy- namic gaussian splatting from causually-captured monocular videos.arXiv preprint arXiv:2406.00434, 2024

    Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy- namic gaussian splatting from causually-captured monocular videos.arXiv preprint arXiv:2406.00434, 2024. 3

    2024 arXiv

  19. [27]

    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. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 13–23, 2023. 2

    2023

  20. [28]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis.arXiv preprint arXiv:2308.09713, 2023

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis.arXiv preprint arXiv:2308.09713, 2023. 1, 2, 5

    2023 arXiv

  21. [29]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InECCV, 2020. 2, 6

    2020

  22. [30]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022. 1, 2

    2022

  23. [31]

    Neural articulated radiance field

    Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 5762–5772, 2021. 1

    2021

  24. [32]

    3d reconstruction of a moving point from a series of 2d projections

    Hyun Soo Park, Takaaki Shiratori, Iain Matthews, and Yaser Sheikh. 3d reconstruction of a moving point from a series of 2d projections. InComputer Vision–ECCV 2010: 11th Eu- ropean Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part...

    2010

  25. [33]

    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 neural radiance fields.ACM Trans. Graph., 40(6), 2021. 5, 6, 8

    2021

  26. [34]

    Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, ...

    2019

  27. [35]

    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. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2, 3, 5, 6

    2021

  28. [36]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Com- puter Vision and Pattern Recognition (CVPR), 2016. 4, 5

    2016

  29. [37]

    Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering

    Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1663...

    2023

  30. [38]

    Monon- erf: Learning a generalizable dynamic radiance field from monocular videos

    Fengrui Tian, Shaoyi Du, and Yueqi Duan. Monon- erf: Learning a generalizable dynamic radiance field from monocular videos. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 17903–17913,

  31. [39]

    Nonrigid structure-from-motion: Estimating shape and mo- tion with hierarchical priors.IEEE transactions on pattern analysis and machine intelligence, 30(5):878–892, 2008

    Lorenzo Torresani, Aaron Hertzmann, and Chris Bregler. Nonrigid structure-from-motion: Estimating shape and mo- tion with hierarchical priors.IEEE transactions on pattern analysis and machine intelligence, 30(5):878–892, 2008. 2

    2008

  32. [40]

    Neural trajectory fields for dynamic novel view syn- thesis.arXiv preprint arXiv:2105.05994, 2021

    Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view syn- thesis.arXiv preprint arXiv:2105.05994, 2021. 1, 2, 3, 4

    2021 arXiv

  33. [41]

    Masked space-time hash encoding for efficient dynamic scene reconstruction.Advances in Neural Informa- tion Processing Systems, 36, 2024

    Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction.Advances in Neural Informa- tion Processing Systems, 36, 2024. 2

    2024

  34. [42]

    Shape of motion: 4d reconstruc- tion from a single video.arXiv preprint arXiv:2407.13764,

    Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruc- tion from a single video.arXiv preprint arXiv:2407.13764,

  35. [43]

    4d gaussian splatting for real-time dynamic scene rendering

    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 rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310–20320, ...

    2024

  36. [44]

    A closed- form solution to non-rigid shape and motion recovery

    Jing Xiao, Jin-xiang Chai, and Takeo Kanade. A closed- form solution to non-rigid shape and motion recovery. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part IV 8, pages 573–587. Springer, 2004. 2

    2004

  37. [45]

    Neural radiance fields for the real world: A survey.arXiv preprint arXiv:2501.13104,

    Wenhui Xiao, Remi Chierchia, Rodrigo Santa Cruz, Xuesong Li, David Ahmedt-Aristizabal, Olivier Salvado, Clinton Fookes, and Leo Lebrat. Neural radiance fields for the real world: A survey.arXiv preprint arXiv:2501.13104,

    arXiv

  38. [46]

    Track anything: Segment anything meets videos.arXiv preprint arXiv:2304.11968, 2023

    Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos.arXiv preprint arXiv:2304.11968, 2023. 4, 5

    2023 arXiv

  39. [47]

    Real-time photorealistic dynamic scene representa- tion and rendering with 4d gaussian splatting.arXiv preprint arXiv:2310.10642, 2023

    Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representa- tion and rendering with 4d gaussian splatting.arXiv preprint arXiv:2310.10642, 2023. 2

    2023 arXiv

  40. [48]

    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. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331–20341, 2024. 1, 2...

    2024

  41. [49]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, 2018. 6

    2018

  42. [50]

    Consistent depth of moving objects in video.ACM Transactions on Graphics (TOG), 40(4):1– 12, 2021

    Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video.ACM Transactions on Graphics (TOG), 40(4):1– 12, 2021. 4

    2021

  43. [51]

    Convolutional sparse cod- ing for trajectory reconstruction.IEEE transactions on pattern analysis and machine intelligence, 37(3):529–540,

    Yingying Zhu and Simon Lucey. Convolutional sparse cod- ing for trajectory reconstruction.IEEE transactions on pattern analysis and machine intelligence, 37(3):529–540,

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.