Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Integrating variational Gaussian Processes into 4D Gaussian Splatting turns deterministic dynamic reconstruction into a probabilistic model that quantifies motion uncertainty, fills unobserved regions, and extrapolates future frames.

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 ↗

T0 review · grok-4.5

2026-07-13 13:40 UTC pith:PTNN3UJZ

load-bearing objection Solid first integration of variational GPs into 4DGS that delivers modest but consistent sparse-view gains, usable uncertainty, and short-horizon extrapolation; the hand-crafted kernel is a real but not fatal limitation. the 2 major comments →

arxiv 2604.02915 v2 pith:PTNN3UJZ submitted 2026-04-03 cs.CV

GP-4DGS: Probabilistic 4D Gaussian Splatting from Monocular Video via Variational Gaussian Processes

Mijeong Kim , Jungtaek Kim , Bohyung Han This is my paper
classification cs.CV
keywords 4D Gaussian SplattingGaussian Processesdynamic novel view synthesisuncertainty quantificationmonocular videovariational inferencemotion extrapolationspatio-temporal kernels
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.

Existing 4D Gaussian Splatting methods reconstruct dynamic scenes from monocular video by deforming Gaussian primitives over time, but treat motion as a fixed, deterministic problem. That leaves them unable to express ambiguity, regularize poorly observed or occluded regions, or predict beyond the training frames. This paper shows that placing a variational Gaussian Process prior over the deformation field, with a composite kernel that separates spatial geometry from temporal periodicity, supplies exactly those missing capabilities. The GP is trained only on high-confidence primitives and then used to guide the rest of the scene, creating a feedback loop that improves reconstruction quality while producing calibrated uncertainty maps. The result is a practical bridge between classical probabilistic modeling and modern neural graphics for monocular dynamic capture.

Core claim

A variational Gaussian Process with a carefully designed spatio-temporal kernel, trained on confident Gaussian primitives and used as a regularizer inside the 4DGS optimization loop, yields higher-fidelity monocular dynamic reconstruction, especially under sparse or occluded observations, while simultaneously supplying free uncertainty estimates and usable temporal extrapolation.

What carries the argument

The composite kernel (anisotropic Matérn on canonical 3D positions plus product of per-axis Matérn and periodic kernels on time) together with the GP-GS alternating optimization: confident primitives train the variational GP; the GP posterior mean then regularizes all primitives.

Load-bearing premise

The hand-designed composite kernel plus a simple confidence threshold on rendering weight are enough to capture the true correlation structure of real non-rigid deformations so that the GP mean helps rather than over-smooths.

What would settle it

On a monocular sequence whose motion is highly non-periodic and multi-object, measure whether adding the GP guidance lowers mPSNR/mLPIPS relative to the identical 4DGS baseline, or whether the rendered uncertainty maps fail to sparsify reconstruction error better than a random baseline (higher AUSE).

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

If this is right

  • Uncertainty maps can be used at inference time to flag unreliable regions for downstream robotics or autonomous systems.
  • Sparse or occluded parts of a monocular video can be completed by propagating motion priors from well-observed primitives without extra sensors or multi-view capture.
  • Future-frame motion can be queried directly from the trained GP without architectural changes or re-training.
  • The same alternating GP-GS loop can be reused as a drop-in probabilistic regularizer for other deformation-based neural graphics pipelines.

Where Pith is reading between the lines

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

  • Because the inducing-point selection already uses time-series embeddings, the same machinery could support online or continual updates when new frames arrive after initial training.
  • The confidence-weighted training set suggests a general pattern: any neural field that can compute per-primitive contribution weights could adopt a similar GP prior for uncertainty-aware regularization.
  • Kernel ablations in the paper imply that replacing the fixed periodic component with a learned spectral mixture could further improve non-periodic long-horizon forecasts.

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

2 major / 5 minor

Summary. GP-4DGS integrates variational Gaussian Processes into 4D Gaussian Splatting for monocular dynamic reconstruction. Deformations of Gaussian primitives are modeled as multi-output GPs with a composite kernel (anisotropic Matérn on canonical space plus a product of per-axis Matérn and periodic kernels on time). Inducing-point variational inference makes the approach tractable for tens of thousands of primitives. An alternating GP-GS loop trains the GP only on high-confidence primitives (selected by cumulative α-blending weight) and then regularizes GS optimization with the GP posterior mean. The claimed benefits are improved novel-view synthesis under sparse observations, calibrated motion uncertainty maps, and short-horizon temporal extrapolation, all demonstrated on DyCheck (full, SoM-5, and challenging subsets) and DA VIS.

Significance. The paper is the first to place a full GP prior over 4DGS deformation fields rather than treating motion as a deterministic optimization problem. If the results hold, it supplies three capabilities that existing 4DGS pipelines lack: (i) principled uncertainty maps that correlate with reconstruction error (AUSE), (ii) usable future-frame prediction without architectural changes, and (iii) observation-adaptive regularization that improves sparse-view geometry. The variational formulation, Chronos-based inducing-point initialization, and the explicit GP-GS alternating schedule are concrete engineering contributions that make the idea practical. Gains on the public DyCheck benchmark, especially the challenging low-overlap subset, and the ablation of the GP-GS loop itself give the claims empirical weight.

major comments (2)
  1. §4.1.2 (Eqs. 10–13) and Algorithm 1: the central claim that the GP supplies an adaptive, data-driven prior rests on the hand-designed composite kernel (spatial Matérn + per-axis Matérn × periodic) together with the hard confidence threshold C_k > τ_C. The manuscript does not test whether this structure is misspecified for non-periodic multi-body or cloth-like motion; if it is, the modest mPSNR/mLPIPS gains on the DyCheck challenging subset (Table 1) and the trajectory smoothing in Figs. 6–7 could simply be over-smoothing rather than learned adaptation. A controlled ablation that replaces the periodic product with a non-periodic Matérn or spectral-mixture kernel on a non-periodic multi-object sequence, or that reports failure cases, is needed to confirm that the posterior mean is not merely a smoothness regularizer.
  2. §5.3.1 / Table 2: future-motion extrapolation is evaluated only against naïve linear extrapolation and only on the last 5/15 frames. Because the temporal kernel already encodes periodicity, the large gap on “periodic motion” scenes is expected by construction; a stronger baseline (e.g., a polynomial or MLP deformation field trained with the same held-out protocol) would better isolate the contribution of the GP. Without it the extrapolation claim remains only partially supported.
minor comments (5)
  1. §4.3 / Algorithm 1: the annealing schedule for τ_δ and the precise values of τ_C, N_GP, λ_GP and the spatial noise std are listed only partially; a single hyper-parameter table would improve reproducibility.
  2. §4.4: Monte-Carlo uncertainty for the 6-D rotation (Eq. 20) is described but the number of samples S and the exact conversion from 6-D to SO(3) are not stated; a short implementation note would help.
  3. Figure 2 and the uncertainty maps: color-scale ranges and whether the maps are normalized per-frame or globally are not specified, making quantitative comparison difficult.
  4. Related Work §2.2: GP-GS [12] is correctly cited for static scenes; a one-sentence clarification that the present work is the first GP prior over time-varying 4DGS deformations would sharpen the novelty claim.
  5. Supplementary Table C: the spectral-mixture kernel slightly outperforms the proposed periodic product on long-horizon extrapolation; a brief discussion of when each kernel is preferred would be useful.

Circularity Check

1 steps flagged

No load-bearing circularity; GP-4DGS is a self-contained engineering integration of variational GPs into 4DGS, evaluated on external held-out benchmarks. Only a minor non-load-bearing self-citation to the authors' prior UA-4DGS appears as an uncertainty baseline.

specific steps
  1. self citation load bearing [Table 3 and §5.3.2 (Uncertainty Quantification)]
    "As shown in Table 3, GP-4DGS consistently outperforms both Random and UA-4DGS [19] baselines. Notably, the performance gap becomes larger when evaluating AUSE on high-quality frames (e.g., top 20 and 40 frames)."

    UA-4DGS is prior work by overlapping authors (Kim & Han). It is cited solely as a comparative baseline for AUSE-MSE, not as a premise that forces the kernel, the variational scheme, or the reconstruction claims. The comparison itself is external (against reconstruction error), so the step is only weakly circular and non-load-bearing for the paper's main results.

full rationale

The paper's derivation chain (composite spatio-temporal kernel in Eqs. 10-13, variational inducing-point ELBO in Eq. 14, alternating GP-GS loop in Algorithm 1, posterior mean/variance for guidance/uncertainty/extrapolation in Eqs. 15-21) is definitional of the proposed method rather than a reduction of a claimed first-principles prediction to its own fitted inputs. Kernel hyperparameters, inducing points Z, and variational parameters are optimized on the confident training subset D_C (C_k > τ_C), but all reported claims are measured against external quantities: masked novel-view metrics on DyCheck (Table 1, including the challenging low-overlap subset), held-out future frames for extrapolation (Table 2 vs. linear baseline), and AUSE sparsification against actual reconstruction error (Table 3). These are not tautological re-statements of the ELBO fit. The sole self-citation (UA-4DGS [19], same first/last authors) is used only as one of several uncertainty baselines in Table 3 and is not invoked to justify uniqueness, force the kernel form, or underwrite any central claim; the method stands independently of it. No self-definitional loops, no uniqueness theorems imported from prior author work, and no ansatz smuggled via citation appear. The hand-designed kernel is an explicit modeling assumption (not circularity). Score 1 reflects only the minor non-load-bearing self-citation; the central results remain independently falsifiable on public data.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 3 invented entities

The central claim rests on standard GP and 4DGS machinery plus a handful of design choices (kernel form, confidence threshold, inducing-point selection, loss weights) that are either learned or set by hand. No new physical entities are postulated; the ‘invented’ pieces are algorithmic constructs whose utility is measured empirically.

free parameters (7)
  • λ_GP = 0.1
    Balancing weight between reconstruction loss and GP guidance loss; fixed at 0.1 without sensitivity study.
  • τ_C (confidence threshold)
    Selects which primitives train the GP; value not numerically reported, only the inequality C_k > τ_C.
  • N_GP (GP update interval) = 2000
    How often the GP is re-trained during GS optimization; set to 2000 iterations.
  • spatial noise std for GP inputs = 0.02
    Gaussian noise injected into canonical positions during GP training; fixed at 0.02.
  • number of inducing points M and M_spatial × M_time factorization
    Controls approximation quality vs. cost; chosen by clustering but exact M not stated in main text.
  • kernel length-scales, variances, periods, Matérn ν = learned after hand init
    All kernel hyper-parameters are optimized via ELBO but initialized by hand (e.g., length-scales 0.001/0.002).
  • τ_δ annealing schedule
    Threshold that decides which primitives receive GP guidance; annealed from start to end values that are not numerically given.
axioms (4)
  • standard math Standard variational sparse GP approximation with inducing points yields a tractable ELBO whose optimum approximates the true GP posterior (Titsias 2009).
    Invoked in §4.2 to justify O(NM²+M³) scaling.
  • domain assumption 4D Gaussian Splatting deformation model (position + 6D rotation) and α-blending rasterizer are correct and differentiable.
    Taken as given from the 4DGS literature (§3.1).
  • ad hoc to paper Cumulative α-blending weight C_k is a reliable proxy for observation confidence of a primitive.
    Defined in Eq. 17 and used to select the GP training set; no external validation that C_k correlates with true geometric reliability.
  • ad hoc to paper A product of per-axis Matérn and periodic kernels adequately models real non-rigid motion correlations for both interpolation and short-horizon extrapolation.
    Kernel form is chosen by the authors (§4.1.2); ablations show it helps but do not prove it is the correct generative model of scene dynamics.
invented entities (3)
  • Composite spatio-temporal kernel (spatial Matérn + temporal Matérn×periodic product) no independent evidence
    purpose: Capture anisotropic geometry and periodic motion so that GP posteriors can regularize unobserved regions and extrapolate.
    Hand-designed; independent evidence is only the empirical ablations in the paper itself.
  • GP-GS alternating optimization loop with confidence-weighted sampling and annealed guidance no independent evidence
    purpose: Jointly refine the GP prior and the 4DGS appearance so that each improves the other.
    Algorithmic construct introduced in Algorithm 1; utility shown only by the paper’s own ablations.
  • Chronos-embedding + k-means inducing-point selection no independent evidence
    purpose: Place inducing points on trajectories that are temporally informative rather than merely spatially dense.
    Specific initialization recipe; Table 4 shows higher ELBO than random/velocity baselines, but no external corroboration.

pith-pipeline@v1.1.0-grok45 · 22687 in / 3372 out tokens · 32617 ms · 2026-07-13T13:40:44.098748+00:00 · methodology

0 comments
read the original abstract

We present GP-4DGS, a novel framework that integrates Gaussian Processes (GPs) into 4D Gaussian Splatting (4DGS) for principled probabilistic modeling of dynamic scenes. While existing 4DGS methods focus on deterministic reconstruction, they are inherently limited in capturing motion ambiguity and lack mechanisms to assess prediction reliability. By leveraging the kernel-based probabilistic nature of GPs, our approach introduces three key capabilities: (i) uncertainty quantification for motion predictions, (ii) motion estimation for unobserved or sparsely sampled regions, and (iii) temporal extrapolation beyond observed training frames. To scale GPs to the large number of Gaussian primitives in 4DGS, we design spatio-temporal kernels that capture the correlation structure of deformation fields and adopt variational Gaussian Processes with inducing points for tractable inference. Our experiments show that GP-4DGS enhances reconstruction quality while providing reliable uncertainty estimates that effectively identify regions of high motion ambiguity. By addressing these challenges, our work takes a meaningful step toward bridging probabilistic modeling and neural graphics.

Figures

Figures reproduced from arXiv: 2604.02915 by Bohyung Han, Jungtaek Kim, Mijeong Kim.

Figure 1
Figure 1. Figure 1: We propose GP-4DGS, a novel integration of Gaussian Processes (GPs) [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Uncertainty quantification. GP-4DGS provides princi [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Motion extrapolation results from GP-4DGS. Our GP [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on the DAVIS dataset under ex [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of inducing points (paper-windmill). In￾ducing points (red) are well-distributed across the canonical space and temporal axis to ensure comprehensive coverage of the scene [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trajectory comparison on the (left) paper-windmill and (right) block scene. GP guidance effectively regularizes motion trajectories, reducing noise and producing physically plausible motion patterns, compared to the baseline approach. 0.0 0.2 0.4 0.6 0.8 1.0 Time (0-1) 1 0 1 2 Deformation (a) Trajectory w/o GP 0.0 0.2 0.4 0.6 0.8 1.0 Time (0-1) 1.5 1.0 0.5 0.0 0.5 1.0 1.5 Deformation (b) Trajectory from GP… view at source ↗
Figure 7
Figure 7. Figure 7: Trajectory comparison on the spin scene between the initial GS reconstruction and the GP guidance in the GP-GS op￾timization. These graphs correspond to the first dimension in 6D rotation. The GP provides accurate and stable motion priors. ability, we adopt the Area Under the Sparsification Error (AUSE), which measures the alignment between estimated uncertainty and actual reconstruction error. As shown in… view at source ↗

Review history (2 revisions) →

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. Future Rendering $\neq$ Future Surface: A Benchmark and Dataset for Dynamic Surface Reconstruction Beyond the Observed Window

    cs.CV 2026-07 conditional novelty 6.0

    FutureSurf, a new benchmark for held-out future surface reconstruction, shows deformation-MLP methods leave a 2-6.6× future-surface gap while rendering quality stays flat.

Reference graph

Works this paper leans on

55 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Mad- dix, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Syndar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Mad- dix, Michael W. Mahoney, Kari Torkkola, Andrew Gor- don Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of tim...

    2024

  2. [2]

    DynaSurfGS: Dynamic surface reconstruction with planar- based Gaussian splatting.arXiv preprint arXiv:2408.13972,

    Weiwei Cai, Weicai Ye, Peng Ye, Tong He, and Tao Chen. DynaSurfGS: Dynamic surface reconstruction with planar- based Gaussian splatting.arXiv preprint arXiv:2408.13972,

    Pith/arXiv arXiv

  3. [3]

    TensoRF: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. InECCV, 2022. 1

    2022

  4. [4]

    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. InICCV, 2021. 2

    2021

  5. [5]

    4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes

    Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wen- zheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In SIGGRAPH, 2024. 2

    2024

  6. [6]

    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 CVPR, 2023. 2

    2023

  7. [7]

    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 ICCV, 2021. 2

    2021

  8. [8]

    Monocular dynamic view synthesis: A reality check

    Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. InNeurIPS, 2022. 6

    2022

  9. [9]

    FastNeRF: High-fidelity neural rendering at 200FPS

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. FastNeRF: High-fidelity neural rendering at 200FPS. InICCV, 2021. 1

    2021

  10. [10]

    GPyTorch: Blackbox matrix- matrix Gaussian process inference with GPU acceleration

    Jacob Gardner, Geoff Pleiss, Kilian Q Weinberger, David Bindel, and Andrew G Wilson. GPyTorch: Blackbox matrix- matrix Gaussian process inference with GPU acceleration. In NIPS, 2018. 6

    2018

  11. [11]

    Motion-aware 3d gaussian splatting for effi- cient dynamic scene reconstruction

    Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for effi- cient dynamic scene reconstruction. InTCSVT, 2024. 2

    2024

  12. [12]

    GP-GS: Gaussian processes for enhanced Gaussian splatting.arXiv preprint arXiv:2502.02283, 2025

    Zhihao Guo, Jingxuan Su, Shenglin Wang, Jinlong Fan, Jing Zhang, Liangxiu Han, and Peng Wang. GP-GS: Gaussian processes for enhanced Gaussian splatting.arXiv preprint arXiv:2502.02283, 2025. 2, 3

    arXiv 2025

  13. [13]

    LRM: Large reconstruction model for single image to 3D

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. InICLR, 2024. 1

    2024

  14. [14]

    Vivid4d: Improving 4d reconstruction from monocular video by video inpainting

    Jiaxin Huang, Sheng Miao, Bangbang Yang, Yuewen Ma, and Yiyi Liao. Vivid4d: Improving 4d reconstruction from monocular video by video inpainting. InICCV, 2025. 6, 7, 12

    2025

  15. [15]

    SC-GS: Sparse-controlled Gaussian splatting for editable dynamic scenes

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. SC-GS: Sparse-controlled Gaussian splatting for editable dynamic scenes. InCVPR,

  16. [16]

    3D Gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. InACM TOG, 2023. 1, 2

    2023

  17. [17]

    3D Gaussian splatting as Markov Chain Monte Carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D Gaussian splatting as Markov Chain Monte Carlo. InNeurIPS, 2024. 2, 3

    2024

  18. [18]

    InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering

    Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering. InCVPR, 2022. 1

    2022

  19. [19]

    4D Gaus- sian splatting in the wild with uncertainty-aware regulariza- tion

    Mijeong Kim, Jongwoo Lim, and Bohyung Han. 4D Gaus- sian splatting in the wild with uncertainty-aware regulariza- tion. InNeurIPS, 2024. 2, 8

    2024

  20. [20]

    DynMF: Neural motion factorization for real-time dynamic view synthesis with 3D Gaussian splatting

    Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. DynMF: Neural motion factorization for real-time dynamic view synthesis with 3D Gaussian splatting. InECCV, 2024. 2

    2024

  21. [21]

    Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds, 2024

    Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds, 2024. arXiv preprint arXiv:2405.17421. 2

    Pith/arXiv arXiv 2024

  22. [22]

    DynIBaR: Neural dynamic image-based rendering

    Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. DynIBaR: Neural dynamic image-based rendering. InCVPR, 2023. 6

    2023

  23. [23]

    Spacetime Gaussian feature splatting for real-time dynamic view syn- thesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime Gaussian feature splatting for real-time dynamic view syn- thesis. InCVPR, 2024. 2, 3, 6

    2024

  24. [24]

    MegaSaM: Accurate, fast and robust structure and motion from casual dynamic videos

    Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holyn- ski, and Noah Snavely. MegaSaM: Accurate, fast and robust structure and motion from casual dynamic videos. InCVPR,

  25. [25]

    Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis

    Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen- Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. InWACV, 2025. 2

    2025

  26. [26]

    Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle

    Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. InCVPR, 2024

    2024

  27. [27]

    Modgs: Dy- namic gaussian splatting from casually-captured monocular videos

    Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lyv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy- namic gaussian splatting from casually-captured monocular videos. InICLR, 2025. 2

    2025

  28. [28]

    VL-Grasp: a 6-Dof interactive grasp policy for language-oriented objects in cluttered indoor scenes

    Yuhao Lu, Yixuan Fan, Beixing Deng, Fangfu Liu, Yali Li, and Shengjin Wang. VL-Grasp: a 6-Dof interactive grasp policy for language-oriented objects in cluttered indoor scenes. InIROS, 2023. 1

    2023

  29. [29]

    3d geometry-aware deformable gaussian splatting for dynamic view synthesis

    Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. InCVPR, 2024. 2

    2024

  30. [30]

    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. 1, 2

    2020

  31. [31]

    DriveWorld: 4D pre-trained scene understanding via world models for au- tonomous driving

    Chen Min, Dawei Zhao, Liang Xiao, Jian Zhao, Xinli Xu, Zheng Zhu, Lei Jin, Jianshu Li, Yulan Guo, Junliang Xing, Liping Jing, Yiming Nie, and Bin Dai. DriveWorld: 4D pre-trained scene understanding via world models for au- tonomous driving. InCVPR, 2024. 1

    2024

  32. [32]

    Instant neural graphics primitives with a multires- olution hash encoding.ACM TOG, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding.ACM TOG, 2022. 1

    2022

  33. [33]

    Barron, Sofien Bouaziz, Dan B Goldman, Steven M

    Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 2

    2021

  34. [34]

    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 TOG, 40(6):1–12, 2021. 6

    2021

  35. [35]

    DeSiRe-GS: 4D street gaussians for static- dynamic decomposition and surface reconstruction for urban driving scenes

    Yuxin Peng et al. DeSiRe-GS: 4D street gaussians for static- dynamic decomposition and surface reconstruction for urban driving scenes. InCVPR, 2025. 1

    2025

  36. [36]

    The 2017 DA VIS challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Ar- bel´aez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 DA VIS challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017. 6

    Pith/arXiv arXiv 2017

  37. [37]

    Gaussian Processes for Machine Learning

    Carl Edward Rasmussen and Christopher KI Williams. Gaussian Processes for Machine Learning. MIT Press, 2006. 1, 2

    2006

  38. [38]

    Rig3DGS: Creating controllable portraits from ca- sual monocular videos.arXiv:2402.03723, 2024

    Alfredo Rivero, ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Rig3DGS: Creating controllable portraits from ca- sual monocular videos.arXiv:2402.03723, 2024. 2

    Pith/arXiv arXiv 2024

  39. [39]

    NeRF- SLAM: Real-time dense monocular SLAM with neural radi- ance fields

    Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF- SLAM: Real-time dense monocular SLAM with neural radi- ance fields. InIROS, 2023. 1

    2023

  40. [40]

    Mod- eling uncertainty for Gaussian splatting.arXiv preprint arXiv:2403.18476, 2024

    Luca Savant, Diego Valsesia, and Enrico Magli. Mod- eling uncertainty for Gaussian splatting.arXiv preprint arXiv:2403.18476, 2024. 2, 8

    Pith/arXiv arXiv 2024

  41. [41]

    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. InCVPR, 2023. 2

    2023

  42. [42]

    Dynamic gaussian marbles for novel view synthesis of casual monocular videos

    Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. InSIGGRAPH, 2024. 2, 6, 7, 13

    2024

  43. [43]

    Variational learning of inducing variables in sparse Gaussian processes

    Michalis K Titsias. Variational learning of inducing variables in sparse Gaussian processes. InAISTATS, 2009. 4

    2009

  44. [44]

    Variational bayes gaussian splatting

    Toon Van de Maele, Ozan Catal, Alexander Tschantz, Christopher L Buckley, and Tim Verbelen. Variational bayes gaussian splatting. InNeurIPS, 2025. 2, 3

    2025

  45. [45]

    D-miso: Editing dynamic 3d scenes using multi-gaussians soup

    Joanna Waczynska, Piotr Borycki, Joanna Kaleta, Slawomir Tadeja, and Przemysław Spurek. D-miso: Editing dynamic 3d scenes using multi-gaussians soup. InNeurIPS, 2024. 2

    2024

  46. [46]

    Shape of Mo- tion: 4D reconstruction from a single video

    Qianqian Wang, Vickie Ye, Hang Gao, Weijia Zeng, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of Mo- tion: 4D reconstruction from a single video. InICCV, 2025. 2, 3, 6, 7, 13

    2025

  47. [47]

    Gflow: Recovering 4D world from monocular video

    Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4D world from monocular video. InAAAI, 2025

    2025

  48. [48]

    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. InCVPR, 2024. 2, 3, 6

    2024

  49. [49]

    Deblur4DGS: 4D Gaussian splatting from blurry monocular video.arXiv preprint arXiv:2412.06424,

    Renlong Wu, Zhilu Zhang, Mingyang Chen, Zifei Yan, and Wangmeng Zuo. Deblur4DGS: 4D Gaussian splatting from blurry monocular video.arXiv preprint arXiv:2412.06424,

    arXiv

  50. [50]

    DrivingSphere: Building a high-fidelity 4D world for closed- loop simulation

    Tianyi Yan, Dongming Wu, Wencheng Han, Junpeng Jiang, Xia Zhou, Kun Zhan, Cheng-zhong Xu, and Jianbing Shen. DrivingSphere: Building a high-fidelity 4D world for closed- loop simulation. InCVPR, 2025. 1

    2025

  51. [51]

    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. In CVPR, 2024. 2, 3, 6

    2024

  52. [52]

    Online learning of neural surface light fields alongside real-time incremental 3D re- construction.IEEE Robotics and Automation Letters, 8(6): 3844–3851, 2023

    Yijun Yuan and Andreas N ¨uchter. Online learning of neural surface light fields alongside real-time incremental 3D re- construction.IEEE Robotics and Automation Letters, 8(6): 3844–3851, 2023. 1

    2023

  53. [53]

    GeoLRM: Geometry-aware large recon- struction model for high-quality 3D Gaussian generation

    Chubin Zhang, Hongliang Song, Yi Wei, Chen Yu, Jiwen Lu, and Yansong Tang. GeoLRM: Geometry-aware large recon- struction model for high-quality 3D Gaussian generation. In NeuIPS, 2024. 1

    2024

  54. [54]

    DrivingGaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. DrivingGaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. InCVPR, 2024. 1

    2024

  55. [55]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. InCVPR, 2019. 4 A. Effect of Time-series Feature Extractor MotivationIn Variational Gaussian Processes (VGPs), the initialization of inducing pointsZis a key factor in summarizing the 4D deformation field. To represent com- plex ...

    2019