pith. sign in

arxiv: 2606.23212 · v1 · pith:I7DSMFGNnew · submitted 2026-06-22 · 💻 cs.CV

Temporally Aware Densification for Dynamic 3D Gaussian Splatting

Vikram Sandu , Mayurdeep Pathak , Rajiv Soundararajan This is my paper

Pith reviewed 2026-06-26 09:04 UTC · model grok-4.3

classification 💻 cs.CV
keywords dynamic 3D Gaussian Splattingdensificationtemporal visibilitydynamic scenes3D reconstructionGaussian splattingmulti-view videocomputer vision
0
0 comments X

The pith

Dynamic 3D Gaussian Splatting can sharpen moving regions by tying densification to each Gaussian's temporal visibility and lifespan.

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

Standard densification in dynamic 3D Gaussian Splatting applies the same rules to all Gaussians even though some appear only briefly during scene motion. Short-lived Gaussians therefore receive little refinement and produce blurry outputs in dynamic areas. The paper introduces a Visibility-Aware Densification framework that derives temporal visibility and lifespan directly from existing training signals to decide when and how much each Gaussian should be split or refined. Two supporting mechanisms adjust the refinement threshold according to lifespan and warp deformation offsets around temporal centers. If the approach holds, dynamic regions improve in quality on standard multi-view benchmarks and the module can be added to other dynamic 3DGS pipelines without extra labels or supervision.

Core claim

By integrating temporal visibility into the densification process, the Visibility-Aware Densification framework ensures Gaussians are refined based on their actual temporal presence. A Temporally-Adaptive Thresholding mechanism adjusts each Gaussian's densification threshold according to its temporal lifespan, and a Temporal Offset Warping design enhances deformation capacity around temporal centers. This approach leads to substantial improvements in the visual quality of dynamic regions and outperforms existing methods on three dynamic multi-view benchmark datasets while generalizing across diverse dynamic 3DGS methods as a plug-and-play component.

What carries the argument

Visibility-Aware Densification (VAD) framework, which computes temporal visibility and lifespan from supervision signals to control densification thresholds and offset warping.

If this is right

  • Dynamic regions receive more balanced refinement because short-lived Gaussians are no longer under-densified.
  • Static regions remain unaffected because their longer lifespans keep original thresholds intact.
  • Highly dynamic Gaussians gain extended deformation capacity through temporal offset warping.
  • The module improves reconstruction when inserted into multiple existing dynamic 3DGS pipelines without retraining from scratch.
  • Overall visual quality rises across three standard dynamic multi-view datasets.

Where Pith is reading between the lines

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

  • Similar temporal signals could be injected into other stages of the 3DGS optimization loop, such as opacity or color updates.
  • The same visibility computation may transfer to non-Gaussian dynamic representations that also rely on sparse supervision over time.
  • Scenes with extreme motion variation would test whether the lifespan-based threshold scaling remains stable or needs further calibration.

Load-bearing premise

Temporal visibility and lifespan can be reliably computed from the existing supervision signals and used to adjust densification thresholds without introducing new artifacts or requiring extra labeled data.

What would settle it

Applying the full VAD pipeline to any of the three benchmark datasets produces no improvement or a measurable drop in PSNR or SSIM specifically on dynamic regions relative to the unmodified baseline method.

Figures

Figures reproduced from arXiv: 2606.23212 by Mayurdeep Pathak, Rajiv Soundararajan, Vikram Sandu.

Figure 1
Figure 1. Figure 1: Left) (A) Existing 3DGS densification fails to densify short-lived dynamic Gaussians. (B) VAD with visibility-weighted gradients and TAT with lifespan-aware thresholds enable densification of short-lived dynamic Gaussians. (C) TOW warps temporal coordinates around center t within window λ t, boosting deformation capacity and densifying highly dynamic Gaussians. Right) Improvement over the densification bas… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed densi￾fication framework. (A) Temporal Offset Warping adaptively warps the input time around each Gaussian’s temporal center, stretching regions near the center and com￾pressing those farther away. (B) We accumu￾late the visibility weighted gradient signal for the N densification frames. (C) We dy￾namically adjust the densification threshold based on temporal scale ψ averaged over … view at source ↗
Figure 3
Figure 3. Figure 3: Temporal Offset Warping (TOW). Top: Warping function W(∆t) vs. time. Uniform mapping has constant unit slope (s = 1), whereas TOW applies a piecewise-linear reparameterization with slope snear > 1 inside the focus window and sfar < 1 outside, expanding near-center off￾sets and compressing distant ones while pre￾serving the total temporal span. Bottom: Signals generated under uniform and warped temporal off… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on Neural 3D Video [13]. Scene: i) cook spinach ii) sear steak iii) cut roasted beef STG Ex4DGS Ours Ground truth [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results on the Interdigital Dataset [24]. Scenes shown top to bottom: i) Birthday, ii) Train: Note that our method preserves finer details in the moving train region. Ours Swift4D STG Ours Ground truth [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on VRU Basketball [34]. Scenes top to bottom: i) DG, ii) GZ [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generalization of the pro￾posed VAD module across diverse baselines [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Occlusion and reappearance. Train scene. Baseline uses only deforma￾tion model, while ours better preserves occluded objects upon reappearance. We revisit the densification strategy in dynamic 3DGS and identify its limita￾tions when modeling fast complex mo￾tion. To address this, we propose a de￾formation formulation complemented by our three key contributions. Together, these modules significantly improve… view at source ↗
Figure 11
Figure 11. Figure 11: Impact of TOW on Gaus￾sian lifespan and Model performance. TOW increases the temporal lifespan of Gaussians, allowing them to remain active over longer motion ranges. This extended lifespan enables more effective densifica￾tion in highly dynamic regions, ultimately yielding higher Masked-PSNR as shown in both sear steak and flame steak scenes. steak scenes. With TOW, Gaussians live longer and accumulate m… view at source ↗
Figure 12
Figure 12. Figure 12: Sudden Appearance. Interdigital Painter scene. STG produces blurry artifacts when objects abruptly enter the camera view, whereas our method preserves finer details in these dynamic regions [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visualization of Gaussians with zero and non-zero deformation coefficients. Gaussians with Mi,f = 0 predominantly correspond to static regions [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison on the Panoptic Sports dataset. Our method better reconstructs fast-moving objects shown in red boxes and improves the overall PSNR. Hyperparameter Analysis of TAT and TOW: We present a hyperparameter analysis of the TOW and TAT modules in [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More qualitative results on the Interdigital Dataset [24]. Scenes top to bottom: i) Painter, ii) Train, iii) Theater. Please note the zoom-in crop of the dynamic regions shown in green box. Ours 4DGaussian SaroGS Ours Ground truth [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: More qualitative comparison on Neural 3D Video [13]. Scene: i) cut roasted beef ii) flame salmon iii) flame steak [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative ablation from the baseline to the full model with the addition of VAD, TAT, and TOW. Each component improves reconstruction in dynamic regions. Scenes shown top to bottom: i) Painter, ii) Theater. 11 Video Comparisons We present video comparisons across the Interdigital [24], Neural 3D Video [16], and VRU Basketball [34] datasets for all evaluated methods. On Interdigital, we compare our appro… view at source ↗
read the original abstract

Despite modeling temporal motion, dynamic 3D Gaussian Splatting (3DGS) methods still inherit a static densification strategy that is ill-suited for dynamic scenes. This neglect of temporal behavior leads to under-reconstructed and blurry dynamic regions, as short-lived Gaussians receive sparse supervision and fail to densify effectively. We propose a Visibility-Aware Densification (VAD) framework that integrates temporal visibility into the densification process, ensuring that Gaussians are refined based on their actual temporal presence. A Temporally-Adaptive Thresholding (TAT) mechanism further adjusts each Gaussian's densification threshold according to its temporal lifespan, promoting balanced refinement of both static and dynamic regions. Finally, a Temporal Offset Warping (TOW) design enhances deformation capacity around temporal centers, extending the lifespan of highly dynamic Gaussians and facilitating more effective densification. Our approach achieves substantial improvements in the visual quality of dynamic regions, outperforming existing methods across three dynamic multi-view benchmark datasets. Moreover, the proposed VAD module generalizes across diverse dynamic 3DGS methods, consistently improving dynamic reconstruction as a plug-and-play component.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper claims that static densification in dynamic 3D Gaussian Splatting leads to under-reconstruction of dynamic regions due to sparse supervision on short-lived Gaussians. It introduces a Visibility-Aware Densification (VAD) framework incorporating Temporally-Adaptive Thresholding (TAT) to modulate thresholds by temporal lifespan and Temporal Offset Warping (TOW) to extend dynamic Gaussian capacity around temporal centers. The approach reportedly yields substantial visual quality gains on three dynamic multi-view benchmarks and generalizes as a plug-and-play module across existing dynamic 3DGS methods.

Significance. If validated, the integration of temporal visibility and lifespan into densification could offer a lightweight, supervision-free improvement to dynamic scene modeling in 3DGS, addressing a practical mismatch between static heuristics and time-varying content without new data requirements.

major comments (2)
  1. [Abstract] Abstract: the central premise that temporal visibility and lifespan can be reliably extracted from existing multi-view supervision signals alone (to drive TAT adjustments) is load-bearing for all three proposed components, yet the abstract provides no mechanism, auxiliary loss, or regularization to stabilize this for short-lived Gaussians that receive sparse gradients; this directly matches the stress-test concern and requires explicit validation in the method section.
  2. [Abstract] Abstract: the generalization claim that VAD improves diverse dynamic 3DGS methods as a plug-and-play component is unsupported by any quantitative cross-method results or ablation tables in the provided text; without such evidence the 'consistent improvement' assertion cannot be assessed.
minor comments (1)
  1. [Abstract] Abstract: the three benchmark datasets are not named; listing them (e.g., D-NeRF, HyperNeRF, etc.) would aid reproducibility assessment.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed feedback on our manuscript. We address each major comment below and commit to revisions that strengthen the presentation of our method and claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central premise that temporal visibility and lifespan can be reliably extracted from existing multi-view supervision signals alone (to drive TAT adjustments) is load-bearing for all three proposed components, yet the abstract provides no mechanism, auxiliary loss, or regularization to stabilize this for short-lived Gaussians that receive sparse gradients; this directly matches the stress-test concern and requires explicit validation in the method section.

    Authors: We agree that the abstract, being a high-level summary, does not detail the extraction mechanism. In the full method section (Section 3), temporal visibility is computed directly from the per-Gaussian contribution to rendered pixels across training views and timesteps using the existing multi-view supervision signals, without auxiliary losses. Lifespan is derived from the temporal extent of non-zero opacity contributions. To address the concern for short-lived Gaussians, we will add explicit validation in the revised manuscript, including a new analysis subsection with gradient flow visualizations and stability metrics on dynamic regions to demonstrate that the TAT modulation remains robust under sparse supervision. revision: yes

  2. Referee: [Abstract] Abstract: the generalization claim that VAD improves diverse dynamic 3DGS methods as a plug-and-play component is unsupported by any quantitative cross-method results or ablation tables in the provided text; without such evidence the 'consistent improvement' assertion cannot be assessed.

    Authors: The referee is correct that the provided manuscript excerpt does not include the supporting quantitative tables. The full paper contains cross-method experiments applying VAD to multiple dynamic 3DGS baselines (e.g., 4DGS, Deformable-3DGS), but we acknowledge these results need to be more prominently featured. We will revise the manuscript to include dedicated quantitative tables and ablation studies demonstrating consistent PSNR/SSIM gains across methods, ensuring the generalization claim is fully substantiated. revision: yes

Circularity Check

0 steps flagged

No circularity: algorithmic modules defined independently of their claimed outputs.

full rationale

The paper introduces VAD, TAT, and TOW as explicit algorithmic components that integrate temporal visibility and lifespan into densification thresholds. No equations, fitted parameters, or self-citations are shown that would make any claimed improvement equivalent to its inputs by construction. The central claims rest on the design of these modules and their empirical performance on external benchmarks, which are independent of the definitions themselves. This is the standard case of a method paper whose derivation chain does not reduce to self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

Review is abstract-only; ledger populated from stated problem framing and proposed components only.

axioms (1)
  • domain assumption Standard 3D Gaussian Splatting representation and optimization assumptions hold for dynamic scenes
    Inherited from prior dynamic 3DGS methods referenced in the abstract.
invented entities (3)
  • Visibility-Aware Densification (VAD) no independent evidence
    purpose: Integrate temporal visibility into the densification process
    New framework introduced to address static densification limitation.
  • Temporally-Adaptive Thresholding (TAT) no independent evidence
    purpose: Adjust each Gaussian's densification threshold by its temporal lifespan
    New mechanism to balance static and dynamic refinement.
  • Temporal Offset Warping (TOW) no independent evidence
    purpose: Enhance deformation capacity around temporal centers to extend lifespan of dynamic Gaussians
    New design element to facilitate densification of highly dynamic elements.

pith-pipeline@v0.9.1-grok · 5735 in / 1354 out tokens · 26923 ms · 2026-06-26T09:04:09.750140+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

46 extracted references · 10 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Barron,J.T.,Mildenhall,B.,Tancik,M.,Hedman,P.,Martin-Brualla,R.,Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5855–5864 (2021)

    2021

  2. [2]

    arXiv preprint arXiv:2501.04782 (2025)

    Bond, A., Wang, J.H., Mai, L., Erdem, E., Erdem, A.: Gaussianvideo: Efficient video representation via hierarchical gaussian splatting. arXiv preprint arXiv:2501.04782 (2025)

    arXiv 2025

  3. [3]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)

    2022

  4. [4]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH) (July 2024)

    Duan,Y.,Wei,F.,Dai,Q.,He,Y.,Chen,W.,Chen,B.:4D-RotorGaussianSplatting: Towards Efficient Novel View Synthesis for Dynamic Scenes. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH) (July 2024)

    2024

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenox- els: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  6. [6]

    In: Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications

    Grubert, G., Barthel, F., Hilsmann, A., Eisert, P.: Improving adaptive density control for 3d gaussian splatting. In: Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. pp. 610–621. SCITEPRESS - Science and Technology Publications (2025). https://doi.org/10.5220/001330850000391...

    work page doi:10.5220/0013308500003912 2025

  7. [7]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH)

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geomet- rically accurate radiance fields. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH). Association for Computing Machinery (2024).https://doi.org/10.1145/3641519.3657428

    work page doi:10.1145/3641519.3657428 2024

  8. [8]

    In: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Huang, H., Zhang, Z., Wu, G., Wang, R.: Pgdgs: Improving few-shot 3d gaussian splatting with progressive gaussian densification. In: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5 (2025).https://doi.org/10.1109/ICASSP49660.2025.10889028

    work page doi:10.1109/icassp49660.2025.10889028 2025

  9. [9]

    ACM Transactions on Graphics (TOG)42(4) (July 2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (TOG)42(4) (July 2023)

    2023

  10. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

    Kim, S., Lee, K., Lee, Y.: Color-cued efficient densification method for 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 775–783 (June 2024)

    2024

  11. [11]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  12. [12]

    In: Proceedings of the Neural Information Processing Systems (2024) 16 V

    Lee, J., Won, C., Jung, H., Bae, I., Jeon, H.G.: Fully explicit dynamic guassian splatting. In: Proceedings of the Neural Information Processing Systems (2024) 16 V. Sandu et al

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3D video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5521–5531 (2022)

    2022

  14. [14]

    In: Proceedings of the ACM Special Interest Group on Com- puter Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia)

    Li, Y., Pang, B., Chen, Y., Wang, G.: Anchored 4d gaussian splatting for dynamic novel view synthesis. In: Proceedings of the ACM Special Interest Group on Com- puter Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia). Association for Computing Machinery, New York, NY, USA (2025).https://doi.org/10. 1145/3757377.3763898,https://doi.org/10.1145/375...

    work page doi:10.1145/3757377.3763898 2025

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime gaussian feature splatting for real-time dynamic view synthesis. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    2024

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    2021

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21136–21145 (2024)

    2024

  18. [18]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)

    2020

  19. [19]

    ACM Trans

    Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG) 41(4), 102:1–102:15 (Jul 2022).https://doi.org/10.1145/3528223.3530127 , https://doi.org/10.1145/3528223.3530127

    work page doi:10.1145/3528223.3530127 2022

  20. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Park, H., Ryu, G., Kim, W.: Dropgaussian: Structural regularization for sparse-view gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Park, J., Bui, M.Q.V., Bello, J.L.G., Moon, J., Oh, J., Kim, M.: Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 26866–26875 (June 2025)

    2025

  22. [22]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia) (2025).https://doi.org/10.1145/3757377.3763993

    Patle, G., Girgaonkar, N., Somraj, N., Soundararajan, R.: AD-GS: Alternating Densification for Sparse-Input 3D Gaussian Splatting. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia) (2025).https://doi.org/10.1145/3757377.3763993

    work page doi:10.1145/3757377.3763993 2025

  23. [23]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    RotaBulò,S.,Porzi,L.,Kontschieder,P.:Revisingdensificationingaussiansplatting. In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    2024

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2017)

    Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., Hog, M., Langlois, T., Gendrot, R., Bureller, O., Schubert, A., Allie, V.: Dataset and pipeline for multi- view light-field video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2017)

    2017

  25. [25]

    In: Proceedings of the European Conference on Computer Vision (ECCV)

    Shaw, R., Nazarczuk, M., Song, J., Moreau, A., Catley-Chandar, S., Dhamo, H., Pérez-Pellitero, E.: Swings: sliding windows for dynamic 3d gaussian splatting. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 37–54. Springer (2024)

    2024

  26. [26]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia) (December 2023).https://doi.org/10.1145/3610548

    Somraj, N., Karanayil, A., Soundararajan, R.: SimpleNeRF: Regularizing sparse in- put neural radiance fields with simpler solutions. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia) (December 2023).https://doi.org/10.1145/3610548. 3618188 Temporally Aware Densification for Dynamic 3D ...

    work page doi:10.1145/3610548 2023

  27. [27]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH) (August 2023).https://doi

    Somraj, N., Soundararajan, R.: ViP-NeRF: Visibility prior for sparse input neural radiance fields. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH) (August 2023).https://doi. org/10.1145/3588432.3591539

    work page doi:10.1145/3588432.3591539 2023

  28. [28]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia)

    Stearns, C., Harley, A.W., Uy, M., Dubost, F., Tombari, F., Wetzstein, G., Guibas, L.: Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques - Asia (SIGGRAPH-Asia). pp. 1–11 (2024)

    2024

  29. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  30. [30]

    Proceedings of the European Conference on Computer Vision (ECCV) (2020)

    Teed, Z., Deng, J.Y.: Raft: Recurrent all-pairs field transforms for optical flow. Proceedings of the European Conference on Computer Vision (ECCV) (2020)

    2020

  31. [31]

    In: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Wang, Z., Wu, G., Wang, Z., Xiao, L., Liu, R., Wu, J., Wang, R.: Hda-gs: Hi- erarchical density-controlled for anisotropic 3d gaussian splatting. In: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). pp. 1–5 (2025).https://doi.org/10.1109/ICASSP49660.2025. 10887967

    work page doi:10.1109/icassp49660.2025 2025

  32. [32]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004).https://doi.org/10.1109/TIP.2003.819861

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004).https://doi.org/10.1109/TIP.2003.819861

    work page doi:10.1109/tip.2003.819861 2004

  33. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20310–20320 (June 2024)

    2024

  34. [34]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2025)

    Wu, J., Peng, R., Wang, Z., Xiao, L., Tang, L., Yan, J., Xiong, K., Wang, R.: Swift4d: Adaptive divide-and-conquer gaussian splatting for compact and efficient reconstruction of dynamic scene. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025)

    2025

  35. [35]

    ACM Transactions on Graphics (TOG)43(6) (November 2024)

    Xu, Z., Xu, Y., Yu, Z., Peng, S., Sun, J., Bao, H., Zhou, X.: Representing long volumetric video with temporal gaussian hierarchy. ACM Transactions on Graphics (TOG)43(6) (November 2024)

    2024

  36. [36]

    In: ACM Multimedia 2024 (2024)

    Yan, J., Peng, R., Tang, L., Wang, R.: 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In: ACM Multimedia 2024 (2024)

    2024

  37. [37]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)

    Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)

    2024

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    2024

  39. [39]

    Yuan, Y., Shen, Q., Yang, X., Wang, X.: 1000+ fps 4d gaussian splatting for dynamic scene rendering (2025),https://arxiv.org/abs/2503.16422

    arXiv 2025

  40. [40]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    Zhang, J., Li, J., Yu, X., Huang, L., Gu, L., Zheng, J., Bai, X.: Cor-gs: Sparse- view 3d gaussian splatting via co-regularization. In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    2024

  41. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 18 V

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effec- tiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 18 V. Sandu et al

    2018

  42. [42]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    Zhang, Z., Hu, W., Lao, Y., He, T., Zhao, H.: Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    2024

  43. [43]

    Zhu, Z., Fan, Z., Jiang, Y., Wang, Z.: Fsgs: Real-time few-shot view synthesis using gaussian splatting. In: Proceedings of the European Conference on Computer Vision (ECCV) (2024) Temporally Aware Densification for Dynamic 3D Gaussian Splatting 19 Supplementary Material The contents of this supplement include

    2024

  44. [44]

    Implementation details

  45. [45]

    Qualitative Comparisons

  46. [46]

    base- line

    Scene-wise quantitative comparisons. 7 Implementation Details We implement our framework in PyTorch, building upon the 3DGS codebase and its differentiable rasterization pipeline. All experiments are conducted on a single NVIDIA A4000 16GB GPU. We employ the Adam optimizer [11] with an initial learning rate of2.6 × 10−4 for Gaussian mean parameters, apply...