Three-dimensional Gaussian Reconstruction of Large-scale Scenes Under Multi-view Geometry Constraints

doi:10.16182/j.issn1004731x.joss.25-0925

Abstract

Abstract:

To enhance the geometry reconstruction quality of the GS algorithm in large-scale scene reconstruction, an optimization method constrained by multi-view geometry reconstruction results was proposed. 2D Gaussian planes were used as geometric primitives to overcome depth anisotropy, and dense depth maps generated by DUSt3R and aligned by sparse point clouds were introduced as constraints. By designing a multi-stage optimization strategy that decouples geometry and rendering, the gradient conflict problem in multi-objective training was solved. Experiments on the MatrixCity dataset indicate that the method surpasses comparison methods in related indicators of geometry reconstruction quality and rendering quality in large-scale scenes. The multi-view reconstruction results demonstrate the effectiveness in improving the geometry reconstruction quality of the GS algorithm.

Key words: Gaussian splatting, large-scale scene reconstruction, 2D Gaussian, multi-view reconstruction, geometry reconstruction quality

CLC Number:

TP391.9

Cui Haohao, Di Yanqiang, Liu Qing, Meng Xianguo. Three-dimensional Gaussian Reconstruction of Large-scale Scenes Under Multi-view Geometry Constraints[J]. Journal of System Simulation, 2026, 38(8): 2353-2363.

Figures/Tables 13

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Table 1

Fig. 6

Table 2

Fig. 7

Table 3

Fig. 8

Table 4

Fig. 9

References 21

[1]	Mildenhall B, Srinivasan P P, Tancik M, et al. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis[J]. Communications of the ACM, 2022, 65(1): 99-106.
[2]	Barron J T, Mildenhall B, Tancik M, et al. Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 5835-5844.
[3]	Müller Thomas, Evans A, Schied C, et al. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding[J]. ACM Transactions on Graphics, 2022, 41(4): 102.
[4]	Wang Peng, Liu Lingjie, Liu Yuan, et al. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction[C]//Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 27171-27183.
[5]	Tancik M, Casser V, Yan Xinchen, et al. Block-NeRF: Scalable Large Scene Neural View Synthesis[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 8238-8248.
[6]	Kerbl Bernhard, Kopanas Georgios, Leimkuehler Thomas, et al. 3D Gaussian Splatting for Real-time Radiance Field Rendering[J]. ACM Transactions on Graphics, 2023, 42(4): 139.
[7]	Guédon Antoine, Lepetit Vincent. SuGaR: Surface-aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-quality Mesh Rendering[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). [S.l. : s.n.], 2024: 5354-5363.
[8]	Yu Mulin, Lu Tao, Xu Linning, et al. GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction[C]//Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 129507-129530.
[9]	Yu Zehao, Sattler Torsten, Geiger Andreas. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes[J]. ACM Transactions on Graphics, 2024, 43(6): 271.
[10]	王锋, 银莹, 王佳炎, 等. 基于高斯泼溅的轻量级重建场景分割方法[J]. 计算机学报, 2025, 48(5): 1232-1243.
	Wang Feng, Yin Ying, Wang Jiayan, et al. Object Segmentation in 3D Reconstructed Scenes Based on Gaussian Splatting[J]. Chinese Journal of Computers, 2025, 48(5): 1232-1243.
[11]	蔡泓锐. 三维重建及其应用中的几何表示研究[D]. 合肥: 中国科学技术大学, 2024.
	Cai Hongrui. Research on Geometric Representations in 3D Reconstruction and Its Applications[D]. Hefei: University of Science and Technology of China, 2024.
[12]	侯佳林, 侯榕榕, 鲍跃全. 基于三维高斯溅射的桥梁自动化三维重建方法[J]. 中国公路学报, 2026, 39(1): 87-98.
	Hou Jialin, Hou Rongrong, Bao Yuequan. Automated 3D Reconstruction of Bridges Based on 3D Gaussian Splatting[J]. China Journal of Highway and Transport, 2026, 39(1): 87-98.
[13]	Lin Jiaqi, Li Zhihao, Tang Xiao, et al. VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 5166-5175.
[14]	Liu Yang, Luo Chuanchen, Fan Lue, et al. CityGaussian: Real-time High-quality Large-scale Scene Rendering with Gaussians[C]//Computer Vision-ECCV 2024. Cham: Springer Nature Switzerland, 2025: 265-282.
[15]	Liu Yang, Luo Chuanchen, Mao Zhongkai, et al. CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-scale Scenes[EB/OL]. (2025-03-04) [2025-05-13]. .
[16]	Huang Binbin, Yu Zehao, Chen Anpei, et al. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields[C]//ACM SIGGRAPH 2024 Conference Papers. New York: ACM, 2024: 32.
[17]	Yang Lihe, Kang Bingyi, Huang Zilong, et al. Depth Anything V2[C]//Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 21875-21911.
[18]	Wang Shuzhe, Leroy Vincent, Cabon Yohann, et al. DUSt3R: Geometric 3D Vision Made Easy[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 20697-20709.
[19]	Gower J C. Generalized Procrustes Analysis[J]. Psychometrika, 1975, 40(1): 33-51.
[20]	Li Yixuan, Jiang Lihan, Xu Linning, et al. MatrixCity: A Large-scale City Dataset for City-scale Neural Rendering and Beyond[C]//2023 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2023: 3182-3192.
[21]	Schönberger Johannes L, Frahm J M. Structure-from-Motion Revisited[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2016: 4104-4113.

方法	MatrixCity-Aerial				MatrixCity-Street
方法	R_PSN	P	R	F₁	R_PSN	P	R	F₁
SuGaR	OOM	OOM	OOM	OOM	19.82	0.053	0.111	0.071
2DGS	21.35	0.207	0.390	0.270	21.50	0.334	0.659	0.443
GOF	17.42	FAIL	FAIL	FAIL	20.32	0.219	0.473	0.300
CityGS	27.46	0.362	0.637	0.462	22.98	0.283	0.689	0.401
CityGS V2	27.23	0.441	0.752	0.556	22.19	0.376	0.759	0.503
本文方法	27.25	0.509	0.767	0.612	22.45	0.419	0.780	0.546

方法	MatrixCity-Aerial			MatrixCity-Street
方法	SSIM	R_PSN	LPIPS	SSIM	R_PSN	LPIPS
SuGaR	OOM	OOM	OOM	0.662	19.82	0.478
2DGS	0.632	21.35	0.562	0.723	21.50	0.477
GOF	0.374	17.42	0.588	0.703	20.32	0.440
CityGS	0.865	27.46	0.204	0.808	22.98	0.301
CityGS V2	0.857	27.23	0.169	0.784	22.19	0.382
本文方法	0.859	27.25	0.174	0.795	22.45	0.341

方法	平均误差
不考虑置信图+Procrustes	0.188 2
考虑置信图+Procrustes	0.119 1
考虑置信图+RANSAC	0.035 0

优化阶段	MatrixCity-Aerial
优化阶段	R_PSN	P	R	F₁
单阶段	27.16	0.508	0.763	0.610
多阶段	27.25	0.509	0.767	0.612