针对过度重建问题的频率强度高斯泼溅算法

doi:10.11996/JG.j.2095-302X.2026030564

摘要/Abstract

摘要：

三维高斯泼溅(3DGS)技术在新视角合成领域表现优异，但其传统的致密化策略在处理细节丰富的复杂场景时，极易引发过度重建(Over-reconstruction)现象，导致少量大体积高斯基元拟合大片区域，渲染图像出现明显模糊并严重丢失高频纹理细节。为克服该问题，提出一种频率强度高斯泼溅算法(FIGS)。首先，从图像局部频域特征出发，定义“频率强度”指标，用于精准量化局部空间内高频细节的丰富程度，并建立对应图像的频率强度图。其次，设计频率强度正则化机制，将渲染图像与真实图像的频率强度图差异引入损失函数，促使算法自适应地向高频细节密集的过度重建区域分配更多的高斯基元。最后，引入基于几何置信度的自适应高斯演化策略，利用多视图一致性先验构建几何置信度因子，通过加权致密化与主动剪枝联合实现高斯演化，动态优化整个场景的高斯分配。在Mip-NeRF360和Tanks&Temples等多个真实基准数据集上的大规模定性与定量实验表明，该算法在结构相似性(SSIM)、学习感知图像块相似度(LPIPS)等指标上取得了更有竞争力的结果，并有效降低了过度重建区域的占比。频率强度高斯泼溅算法能够有效缓解重建过程中的模糊，科学合理地调度高斯基元分布，在新视角合成的细节保真度和整体视觉质量上表现优异，为复杂真实场景的高精度三维重建提供了新的解决思路。

关键词: 新视角合成, 3D高斯泼溅, 三维重建, 过度重建, 频域分析

Abstract:

3D Gaussian Splatting (3DGS) technology has demonstrated exceptional performance in the field of novel view synthesis. However, when processing complex scenes with abundant details, its conventional densification strategy is highly susceptible to the over-reconstruction phenomenon. This leads to a small number of large-volume Gaussian primitives fitting extensive areas, resulting in noticeable blurriness in rendered images and a severe loss of high-frequency texture details. To overcome this issue, a novel Frequency Intensity Gaussian Splatting (FIGS) algorithm was proposed. First, starting from local image frequency-domain characteristics, a “Frequency Intensity” metric was defined to precisely quantify the richness of high-frequency details within local spatial regions, establishing corresponding frequency intensity maps. Second, a frequency-intensity regularization mechanism was designed. By incorporating the discrepancy between the frequency-intensity maps of the rendered images and the ground truth into the loss function, this mechanism drove the algorithm to adaptively allocate more Gaussian primitives to over-reconstructed regions densely populated with high-frequency details. Finally, a confidence-based adaptive Gaussian evolution strategy was introduced. Leveraging multi-view consistency priors to construct a geometric confidence factor, this strategy achieved robust Gaussian evolution through a joint mechanism of weighted densification and active pruning, thereby dynamically optimizing Gaussian allocation across the entire scene. Extensive qualitative and quantitative experiments conducted on multiple real-world benchmark datasets, including Mip-NeRF360 and Tanks & Temples, demonstrated that the proposed algorithm achieved highly competitive results in evaluation metrics such as Structural Similarity (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS), effectively reducing the proportion of over-reconstructed regions. The Frequency Intensity Gaussian Splatting algorithm could effectively alleviate blurriness during the reconstruction process and scientifically orchestrate the spatial distribution of Gaussian primitives. It exhibited outstanding performance in detail fidelity and overall visual quality for novel view synthesis, providing a robust new solution for high-precision 3D reconstruction of complex real-world scenes.

Key words: novel view synthesis, 3D Gaussian splatting, 3D reconstruction, over-reconstruction, frequency-domain analysis

中图分类号:

TP391.41

廖健康, 张严辞. 针对过度重建问题的频率强度高斯泼溅算法[J]. 图学学报, 2026, 47(3): 564-575.

LIAO Jiankang, ZHANG Yanci. Frequency intensity Gaussian splatting for over-reconstruction issue[J]. Journal of Graphics, 2026, 47(3): 564-575.

图/表 13

图1 3D高斯中的过度重建((a) 渲染图像；(b) 真实图像；(c) 高斯椭球可视化)

Fig. 1 Over-reconstruction in 3D Gaussian Splatting ((a) Rendered image; (b) Ground truth; (c) Visualization of Gaussian ellipsoids)

图2 FIGS算法流程

Fig. 2 Algorithm pipeline of FIGS

图3 频率强度图生成算法流程(N=5)

Fig. 3 Algorithm pipeline for frequency intensity map generation (N=5)

图4 稀疏视角的错误几何结构

Fig. 4 Incorrect geometric structure from sparse views

表1 合成质量定量实验结果

Table 1 Quantitative results of rendering quality

Method	Mip-NeRF360			Tanks & Temples			Deep Blending			Ref-Real
Method	SSIM↑	PSNR↑	LPIPS↓	SSIM↑	PSNR↑	LPIPS↓	SSIM↑	PSNR↑	LPIPS↓	SSIM↑	PSNR↑	LPIPS↓
3DGS	0.815	27.21	0.214	0.841	23.14	0.183	0.903	29.41	0.243	0.636	23.52	0.277
SteepGS*	0.809	27.31	0.221	0.845	23.55	0.178	0.904	29.62	0.246	0.642	23.60	0.269
Perceptual-GS*	0.829	27.74	0.185	0.856	23.87	0.150	0.906	29.88	0.231	0.643	23.62	0.249
FreGS	0.826	27.85	0.209	0.849	23.96	0.178	0.904	29.93	0.240	一	一	一
Pixel-GS*	0.818	27.48	0.207	0.846	23.74	0.175	0.899	29.47	0.249	0.642	23.63	0.264
AbsGS*	0.820	27.49	0.191	0.854	23.66	0.156	0.898	29.55	0.235	0.634	23.39	0.269
Ours	0.827	27.57	0.174	0.857	23.64	0.137	0.898	29.26	0.231	0.648	23.59	0.236

图5 合成质量对比

Fig. 5 Comparison of rendering quality

图6 高斯点云可视化

Fig. 6 Visualization of gaussian point clouds

图7 依据频率强度识别过度重建区域 ((a) 渲染图像及局部放大；(b) 识别出的区域)

Fig. 7 Identifying over-reconstruction regions based on frequency intensity ((a) Rendered images with local zoom-ins; (b) Identified regions)

表2 频率强度差距过大区域所占图像百分比

Table 2 Percentage of regions with large frequency intensity differences in images

Method	Mip-NeRF360 (outdoor)	Tanks & Tem-ples	Ref-Real
3DGS	9.65	2.89	5.34
Pixel-GS	6.74	2.45	4.43
Perceptual-GS	4.39	1.93	3.07
AbsGS	4.07	1.52	3.15
Ours	1.80	0.86	2.00

表3 各方法分配高斯球数量百分比

Table 3 The percentage of Gaussians allocated by each method

Method	Mip-NeRF360 (outdoor)	Tanks & Tem-ples	Ref-Real
3DGS	1.83	1.72	2.66
Pixel-GS	2.75	1.43	3.30
Perceptual-GS	3.11	1.83	4.06
AbsGS	2.53	2.02	3.40
Ours	4.00	2.20	5.50

图8 频率强度图及频率强度正则化过程可视化((a) 频率强度图可视化；(b) 频率强度正则化过程可视化)

Fig. 8 Visualization of frequency intensity maps and the frequency-intensity regularization process((a) visualization of frequency intensity maps; (b) visualization of the frequency-intensity regularization process)

表4 消融实验定量分析结果

Table 4 Ablation study quantitative results

Method	Mip-NeRF360			Tanks & Temples
Method	SSIM↑	PSNR↑	LPIPS↓	SSIM↑	PSNR↑	LPIPS↓
Base	0.815	27.21	0.214	0.841	23.14	0.183
Base+FI	0.817	27.26	0.190	0.853	23.66	0.157
Base+GE	0.818	27.48	0.192	0.857	23.72	0.148
Base+FI+GE	0.827	27.57	0.174	0.857	23.64	0.137

图9 FIGS各模块的视觉消融实验

Fig. 9 Ablation study on the effectiveness of FIGS components

参考文献 43

[1]	KERBL B, KOPANAS G, LEIMKÜHLER T, et al. 3D Gaussian splatting for real-time radiance field rendering[J]. ACM Transactions on Graphics (TOG), 2023, 42(4): 139.
[2]	MILDENHALL B, SRINIVASAN P P, TANCIK M, et al. NeRF: representing scenes as neural radiance fields for view synthesis[J]. Communications of the ACM, 2021, 65(1): 99-106.
[3]	SCHÖNBERGER J L, FRAHM J M. Structure-from-motion revisited[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2016: 4104-4113.
[4]	DU Y L, ZHANG Y N, YU H X, et al. Neural radiance flow for 4d view synthesis and video processing[C]//2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 14304-14314.
[5]	LI Z Q, NIKLAUS S, SNAVELY N, et al. Neural scene flow fields for space-time view synthesis of dynamic scenes[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 6494-6504.
[6]	LI T Y, SLAVCHEVA M, ZOLLHOEFER M, et al. Neural 3D video synthesis from multi-view video[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 5511-5521.
[7]	LIANG J B, ZHANG L Q, ZHAO Z R, et al. InfNeRF: towards Infinite scale NeRF rendering with O(log n) space complexity[C]//SIGGRAPH Asia 2024 Conference Papers. New York: ACM, 2024: 1.
[8]	TANCIK M, CASSER V, YAN X C, et al. Block-NeRF: scalable large scene neural view synthesis[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 8238-8248.
[9]	MI Z X, XU D. Switch-NeRF: learning scene decomposition with mixture of experts for large-scale neural radiance fields[EB/OL]. (2023-02-02) [2025-02-27]. https://openreview.net/forum?id=PQ2zoIZqvm.
[10]	OECHSLE M, PENG S Y, GEIGER A. UNISURF: unifying neural implicit surfaces and radiance fields for multi-view reconstruction[C]//2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 5569-5579.
[11]	WANG P, LIU L J, LIU Y, et al. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction[C]//The 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 2081.
[12]	YARIV L, GU J T, KASTEN Y, et al. Volume rendering of neural implicit surfaces[C]//The 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 367.
[13]	REISER C, PENG S Y, LIAO Y Y, et al. KiloNeRF: speeding up neural radiance fields with thousands of tiny MLPs[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 14315-14325.
[14]	WANG H, REN J, HUANG Z, et al. R2L: distilling neural Radiance field to neural light field for efficient novel view synthesis[C]//The 17th European Conference on Computer Vision. Cham: Springer, 2022: 612-629.
[15]	NEFF T, STADLBAUER P, PARGER M, et al. DONeRF: towards real‐time rendering of compact neural radiance fields using depth oracle networks[J]. Computer Graphics Forum, 2021, 40(4): 45-59.
[16]	FRIDOVICH-KEIL S, YU A, TANCIK M, et al. Plenoxels: radiance fields without neural networks[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 5491-5500.
[17]	CHEN A P, XU Z X, GEIGER A, et al. TensoRF: tensorial radiance fields[C]//The 17th European Conference on Computer Vision. Cham: Springer, 2022: 333-350.
[18]	DENG K L, LIU A, ZHU J Y, et al. Depth-supervised NeRF: fewer views and faster training for free[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 12872-12881.
[19]	XIONG H L. SparseGS: real-time 360° sparse view synthesis using Gaussian splatting[D]. Los Angeles: University of California, 2024.
[20]	ZHU Z H, FAN Z W, JIANG Y F, et al. FSGS: real-time few-shot view synthesis using Gaussian splatting[C]//The 18th European Conference on Computer Vision. Cham: Springer, 2024: 145-163.
[21]	CHEN Y D, XU H F, ZHENG C X, et al. MVSplat: efficient 3D Gaussian splatting from sparse multi-view images[C]//The 18th European Conference on Computer Vision. Cham: Springer, 2024: 370-386.
[22]	FAN Z W, WANG K, WEN K R, et al. LightGaussian: unbounded 3D Gaussian compression with 15x reduction and 200+ FPS[C]//The 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 4447.
[23]	LEE J C, RHO D, SUN X Y, et al. Compact 3D Gaussian representation for radiance field[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 21719-21728.
[24]	NIEDERMAYR S, STUMPFEGGER J, WESTERMANN R. Compressed 3D Gaussian splatting for accelerated novel view synthesis[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 10349-10358.
[25]	WANG P H, WANG Y H, WANG D L, et al. Steepest descent density control for compact 3D Gaussian splatting[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2025: 26663-26672.
[26]	LIU Y, LUO C C, FAN L, et al. CityGaussian: real-time high-quality large-scale scene rendering with Gaussians[C]// The 18th European Conference on Computer Vision. Cham: Springer, 2024: 265-282.
[27]	KERBL B, MEULEMAN A, KOPANAS G, et al. A hierarchical 3D Gaussian representation for real-time rendering of very large datasets[J]. ACM Transactions on Graphics (TOG), 2024, 43(4): 62.
[28]	LIN J Q, LI Z H, TANG X, et al. VastGaussian: vast 3D Gaussians for large scene reconstruction[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 5166-5175.
[29]	YANG Z Y, GAO X Y, SUN Y T, et al. Spec-Gaussian: anisotropic view-dependent appearance for 3D Gaussian splatting[C]//The 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 1956.
[30]	LU T, YU M L, XU L N, et al. Scaffold-GS: Structured 3D Gaussians for view-adaptive rendering[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 20654-20664.
[31]	YU Z H, CHEN A P, HUANG B B, et al. Mip-splatting: Alias-free 3D Gaussian splatting[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 19447-19456.
[32]	YE Z X, LI W Y, LIU S D, et al. AbsGS: recovering fine details in 3D Gaussian splatting[C]//The 32nd ACM International Conference on Multimedia. New York: ACM, 2024: 1053-1061.
[33]	YU Z H, SATTLER T, GEIGER A. Gaussian opacity fields: efficient adaptive surface reconstruction in unbounded scenes[J]. ACM Transactions on Graphics (TOG), 2024, 43(6): 271.
[34]	ZHANG Z, HU W B, LAO Y X, et al. Pixel-GS: density control with pixel-aware gradient for 3D Gaussian splatting[C]//The 18th European Conference on Computer Vision. Cham: Springer, 2024: 326-342.
[35]	ZHOU H B, NI Z K. Perceptual-GS: scene-adaptive perceptual densification for Gaussian splatting[EB/OL]. (2025-06-14) [2025-06-20]. https://arxiv.org/abs/2506.12400.
[36]	FANG G C, WANG B. Mini-splatting: representing scenes with a constrained number of Gaussians[C]The 18th European Conference on Computer Vision. Cham: Springer, 2024: 165-181.
[37]	ZHANG J H, ZHAN F N, XU M Y, et al. FreGS: 3D Gaussian splatting with progressive frequency regularization[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 21424-21433.
[38]	CHENG K, LONG X X, YANG K Z, et al. GaussianPro:3D Gaussian splatting with progressive propagation[EB/OL]. (2024-05-02) [2025-06-25]. https://openreview.net/forum?id=lQ3SEBH1gF.
[39]	BARANOWSKI B, ESPOSITO S, GSCHOßMANN P, et al. ConeGS: error-guided densification using pixel cones for improved reconstruction with fewer primitives[EB/OL]. (2025-11-10) [2025-11-21]. https://arxiv.org/abs/2511.06810.
[40]	BARRON J T, MILDENHALL B, VERBIN D, et al. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 5460-5469.
[41]	KNAPITSCH A, PARK J, ZHOU Q Y, et al. Tanks and temples: benchmarking large-scale scene reconstruction[J]. ACM Transactions on Graphics (TOG), 2017, 36(4): 78.
[42]	HEDMAN P, PHILIP J, PRICE T, et al. Deep blending for free-viewpoint image-based rendering[J]. ACM Transactions on Graphics (TOG), 2018, 37(6): 257.
[43]	VERBIN D, HEDMAN P, MILDENHALL B, et al. Ref-NeRF: structured view-dependent appearance for neural radiance fields[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(11): 9426-9437. DOI URL