CAFe-GS: Compactness-Aware Frequency-Guided Densification for 3D Gaussian Splatting
CAFe-GS: Densification orientée compacité par guidage fréquentiel pour le 3D Gaussian Splatting
Résumé
3D Gaussian Splatting (3DGS) represents scenes using Gaussian primitives and enables real-time novel view synthesis. Adaptive Density Control (ADC), a key part of the pipeline, governs when to densify these primitives to balance reconstruction quality and efficiency. In the original 3DGS pipeline, densification is triggered by a thresholded positional-gradient criterion. However, this criterion frequently selects already well-covered regions, leading to redundant primitives and providing weak control over the balance between reconstruction quality and compactness (i.e., fidelity versus primitive count). In CAFe-GS, we propose a new densification criterion based on a per-Gaussian score obtained by mapping per-pixel rendering errors back to the contributing primitives, using their effective-opacity under front-to-back alpha compositing as weights. The score is then modulated by frequency guidance derived from Laplacian-of-Gaussian responses, promoting detail-rich, high-frequency areas in contrast to smooth or already well-reconstructed regions. This criterion drives densification through standard cloning and splitting operations. CAFe-GS provides a clearer, single-parameter handle on the quality–compactness balance. Experiments on standard benchmarks show that CAFe-GS achieves comparable PSNR using ≈2–4× fewer Gaussians at matched quality, and up to 12–15× fewer Gaussians at a controlled PSNR trade-off.
Le 3D Gaussian Splatting (3DGS) représente une scène par un modèle de gaussiennes 3D et permet la synthèse de vues nouvelles en temps réel. Un élément clé du pipeline d'optimisation est le contrôle adaptatif de densité (Adaptive Density Control, ADC), qui sélectionne les gaussiennes à densifier (c.-à-d. dupliquer) afin d’augmenter la capacité du modèle. Dans 3DGS, cette densification tend à sélectionner des zones déjà bien reconstruites, entraînant des duplications redondantes et un contrôle limité du compromis qualité–compacité (fidélité vs. nombre de primitives). Nous proposons CAFe-GS, une stratégie de densification orientée compacité, qui utilise (i) un score d’erreur par gaussienne, obtenu en redistribuant l’erreur de rendu vers les gaussiennes, et (ii) un guidage fréquentiel favorisant la densification dans les régions riches en détails tout en pénalisant les zones lisses. Un paramètre de seuil unique permet un contrôle interprétable sur le compromis. CAFe-GS atteint une qualité comparable en PSNR à 3DGS avec 2–4× moins de gaussiennes, et jusqu’à 12–15× moins avec une baisse de qualité contrôlée.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
![]()
Cite 10.1109/iccv48922.2021.00580 Autre Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., & Srinivasan, P. P. (2021). Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 5835–5844). IEEE. 2021 IEEE/CVF International Conference on Computer Vision (ICCV). https://doi.org/10.1109/iccv48922.2021.00580
![]()
Cite 10.1007/978-3-031-73036-8_20 Autre Rota Bulò, S., Porzi, L., & Kontschieder, P. (2024). Revising Densification in Gaussian Splatting. In Lecture Notes in Computer Science (pp. 347–362). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-73036-8_20
![]()
Cite 10.48550/arXiv.2411.10133 Article Deng, X., Diao, C., Li, M., Yu, R., & Xu, D. (2024). Efficient Density Control for 3D Gaussian Splatting (Version 4). arXiv. https://doi.org/10.48550/ARXIV.2411.10133
![]()
Cite 10.48550/arXiv.2508.12313 Article Deng, X., Diao, C., Li, M., Yu, R., & Xu, D. (2025). Improving Densification in 3D Gaussian Splatting for High-Fidelity Rendering (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2508.12313
![]()
Cite 10.48550/arXiv.2406.07499 Article Fan, L., Yang, Y., Li, M., Li, H., & Zhang, Z. (2024). Trim 3D Gaussian Splatting for Accurate Geometry Representation (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2406.07499
![]()
Cite 10.52202/079017-4447 Autre Fan, Z., Wang, K., Wang, Z., Wen, K., Xu, D., & Zhu, Z. (2024). LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS. In Advances in Neural Information Processing Systems 37 (pp. 140138–140158). Neural Information Processing Systems Foundation, Inc. (NeurIPS). Advances in Neural Information Processing Systems 37. https://doi.org/10.52202/079017-4447
![]()
Cite 10.1007/978-3-031-72980-5_10 Autre Fang, G., & Wang, B. (2024). Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians. In Lecture Notes in Computer Science (pp. 165–181). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-72980-5_10
![]()
Cite 10.1109/cvpr52734.2025.00558 Autre Hanson, A., Tu, A., Singla, V., Jayawardhana, M., Zwicker, M., & Goldstein, T. (2025). PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 5949–5958). IEEE. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). https://doi.org/10.1109/cvpr52734.2025.00558
![]()
Cite hal-01886237 Article Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, et al.. Deep Blending for Free-Viewpoint Image-Based Rendering. ACM Transactions on Graphics, 2018, 37 (6), ⟨10.1145/3272127.3275084⟩. ⟨hal-01886237⟩
![]()
Cite hal-04088161 Article Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 2023, 42 (4), pp.1-14. ⟨10.1145/3592433⟩. ⟨hal-04088161v2⟩
![]()
Cite 10.1109/cvprw63382.2024.00082 Autre Kim, S., Lee, K., & Lee, Y. (2024). Color-cued Efficient Densification Method for 3D Gaussian Splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (pp. 775–783). IEEE. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). https://doi.org/10.1109/cvprw63382.2024.00082
![]()
Cite 10.1145/3072959.3073599 Autre Knapitsch, A., Park, J., Zhou, Q.-Y., & Koltun, V. (2017). Tanks and temples. ACM Transactions on Graphics, 36(4), 1–13. https://doi.org/10.1145/3072959.3073599
![]()
Cite 10.1145/3680528.3687694 Autre Mallick, S. S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F. V., & De La Torre, F. (2024). Taming 3DGS: High-Quality Radiance Fields with Limited Resources. In SIGGRAPH Asia 2024 Conference Papers (pp. 1–11). ACM. SA ’24: SIGGRAPH Asia 2024 Conference Papers. https://doi.org/10.1145/3680528.3687694
![]()
Cite hal-04550892 Objet présenté à une conférence Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, George Drettakis. Reducing the Memory Footprint of 3D Gaussian Splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, May 2024, Philadelphia, United States. ⟨10.1145/3651282⟩. ⟨hal-04550892v2⟩
![]()
Cite hal-05444229 Objet présenté à une conférence Stéphane Pateux, Matthieu Gendrin, Luce Morin, Théo Ladune, Xiaoran Jiang. BOGausS: Better Optimized Gaussian Splatting. 2025 33rd European Signal Processing Conference (EUSIPCO), Sep 2025, Palermo, Italy. pp.765-769, ⟨10.23919/EUSIPCO63237.2025.11226626⟩. ⟨hal-05444229⟩
![]()
Cite 10.48550/arXiv.2409.06765 Article Ye, V., Li, R., Kerr, J., Turkulainen, M., Yi, B., Pan, Z., Seiskari, O., Ye, J., Hu, J., Tancik, M., & Kanazawa, A. (2024). gsplat: An Open-Source Library for Gaussian Splatting (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2409.06765
![]()
Cite 10.1109/cvpr52734.2025.02484 Autre Zhang, Y., Jia, W., Niu, W., & Yin, M. (2025). GaussianSpa: An ‘Optimizing-Sparsifying’ Simplification Framework for Compact and High-Quality 3D Gaussian Splatting. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 26673–26682). IEEE. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). https://doi.org/10.1109/cvpr52734.2025.02484