An adaptive-pruning 3D Gaussian SLAM method for real-time high-fidelity indoor scene modeling

doi:10.11947/j.AGCS.2026.20260050

Abstract

Abstract:

In recent years, robotics and augmented reality (AR) applications have imposed increasingly stringent requirements on the real-time performance and interpretability of dense SLAM. As an explicit 3D scene representation, 3D Gaussian splatting (3DGS) combines fast rendering with structural interpretability and has emerged as a promising direction for high-quality mapping. However, existing 3DGS-SLAM systems still struggle to simultaneously achieve high reconstruction accuracy and real-time efficiency in online mapping, mainly due to the lack of adaptive control over the number of Gaussian primitives. To address this issue, we propose an adaptive-pruning 3D Gaussian SLAM method for real-time, high-fidelity modeling. The proposed approach explicitly incorporates a “map scale/Gaussian budget” term into the mapping optimization objective and introduces a score-driven Gaussian management mechanism. Specifically, during optimization, multi-view photometric consistency residuals, occlusion awareness, and viewpoint coverage are jointly leveraged to online estimate the contribution of each Gaussian to geometric consistency and appearance fidelity. Based on these scores, adaptive densification and priority-based pruning are performed in a unified manner to dynamically regulate model complexity. In addition, a learnable mask coupled with a sparsity regularization term is introduced to promote the adaptive contraction of the effective Gaussian set, thereby reducing GPU memory usage and rendering overhead. Experiments on public benchmark datasets demonstrate that the proposed method maintains stable reconstruction quality even at an approximately 65% pruning ratio, while substantially reducing computational and storage costs. These results indicate that the proposed approach improves online mapping efficiency without compromising map visualization quality, providing an effective scale-adaptive control scheme for real-time, high-fidelity 3DGS-SLAM mapping.

Key words: 3DGS, SLAM, real-time mapping, adaptive pruning, visualization

CLC Number:

P237

Jun Gong, Jianjun Luo, Shengnan Ke, Shibin Li, Weicong Chen, Le Qin, Shengjun Tang. An adaptive-pruning 3D Gaussian SLAM method for real-time high-fidelity indoor scene modeling[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(7): 1266-1277.

Add to citation manager EndNote|Reference Manager|ProCite|BibTeX|RefWorks

URL: http://xb.chinasmp.com/EN/10.11947/j.AGCS.2026.20260050

http://xb.chinasmp.com/EN/Y2026/V55/I7/1266

Figures/Tables 16

Fig. 1

Tab. 1

Tab. 2

Fig. 2

Tab. 3

Tab. 4

Tab. 5

Fig. 3

Tab. 6

Tab. 7

Tab. 8

Tab. 9

Tab. 10

Fig. 4

Fig. 5

Tab. 11

References 30

[1]	张继贤, 刘飞. 视觉SLAM环境感知技术现状与智能化测绘应用展望[J]. 测绘学报, 2023, 52(10): 1617-1630. DOI: . doi: 10.11947/j.AGCS.2023.20220240
	Zhang Jixian, Liu Fei. Review of visual SLAM environment perception technology and intelligent surveying and mapping application[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(10): 1617-1630. DOI: . doi: 10.11947/j.AGCS.2023.20220240
[2]	Mur-Artal R, Montiel J M M, Tardos J D. ORB-SLAM: a versatile and accurate monocular SLAM system[J]. IEEE transactions on robotics, 2015, 31(5): 1147-1163.
[3]	Newcombe R A, Fox D, Seitz S M. DynamicFusion: reconstruction and tracking of non-rigid scenes in real-time[C]//Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston: IEEE, 2015: 343-352.
[4]	刘佳, 张增伟, 陈大鹏. 基于VSLAM的室内场景重建与虚实遮挡的边缘优化方法[J]. 计算机辅助设计与图形学学报, 2025, 37(5): 744-752.
	Liu Jia, Zhang Zengwei, Chen Dapeng. The method of edge optimization for indoor scene reconstruction and virtual-real occlusion based on VSLAM[J]. Journal of Computer-Aided Design & Computer Graphics, 2025, 37(5): 744-752.
[5]	马威, 涂强, 潘建平, 等. 桥梁实景三维高斯辐射场建模[J]. 测绘学报, 2024, 53(9): 1694-1705. DOI: . doi: 10.11947/j.AGCS.2024.20240071
	Ma Wei, Tu Qiang, Pan Jianping, et al. 3D Gaussian radiation field modeling for real-scene bridges[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(9): 1694-1705. DOI: . doi: 10.11947/j.AGCS.2024.20240071
[6]	谭臻, 牛中颜, 张津浦, 等. SLAM新机遇—高斯溅射技术[J]. 中国图象图形学报, 2025, 30(6): 1792-1807.
	Tan Zhen, Niu Zhongyan, Zhang Jinpu, et al. New opportunities in SLAM—Gaussian splatting technology[J]. Journal of Image and Graphics, 2025, 30(6): 1792-1807.
[7]	Yan Chi, Qu Delin, Xu Dan, et al. GS-SLAM: dense visual SLAM with 3D Gaussian splatting[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 19595-19604.
[8]	Keetha N, Karhade J, Jatavallabhula K M, et al. SplaTAM: splat, track & map 3D Gaussians for dense RGB-D SLAM[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 21357-21366.
[9]	Matsuki H, Murai R, Kelly P H J, et al. Gaussian splatting SLAM[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 18039-18048.
[10]	Ha S, Yeon J, Yu H. RGBD GS-ICP SLAM[C]//Proceedings of 2024 European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 180-197.
[11]	Huang Huajian, Li Longwei, Cheng Hui, et al. Photo-SLAM: real-time simultaneous localization and photorealistic mapping for monocular, stereo, and RGB-D cameras[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 21584-21593.
[12]	Liu Yang, Luo Chuanchen, Fan Lue, et al. CityGaussian: real-time high-quality large-scale scene rendering with Gaussians[C]//Proceedings of 2024 European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 265-282.
[13]	Kerbl B, Kopanas G, Leimkühler T, et al. 3D Gaussian splatting for real-time radiance field rendering[J]. ACM Transactions on Graphics, 2023, 42(4): 139:1-139:14.
[14]	Liu Shiyong, Tang Xiao, Li Zhihao, et al. OccluGaussian: occlusion-aware Gaussian splatting for large scene reconstruction and rendering[C]//Proceedings of 2025 IEEE/CVF International Conference on Computer Vision. Honolulu: IEEE, 2025: 26643-26652.
[15]	Deng Tianchen, Chen Yaohui, Yang Jianfei, et al. CGS-SLAM: compact 3D Gaussian splatting for dense visual SLAM[C]//Proceedings of 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems. Hangzhou: IEEE, 2025: 1606-1613.
[16]	Zhao Jie, Zhu Yongjian, Zhao Ruodeng. Large-scale compact 3D Gaussian splatting SLAM[J]. Applied Optics, 2025, 64(31): 9185-9195.
[17]	Zhang Zhaoliang, Song Tianchen, Lee Y, et al. LP-3DGS: learning to prune 3D Gaussian splatting[J]. Advances in Neural Information Processing Systems, 2024, 37: 122434-122457.
[18]	Hanson A, Tu A, Singla V, et al. PUP 3D-GS: principled uncertainty pruning for 3D Gaussian splatting[C]//Proceedings of 2025 Computer Vision and Pattern Recognition Conference. Nashville: IEEE, 2025: 5949-5958.
[19]	钱虞杰, 丁丹丹. 基于帧间自适应预测的动态点云压缩[J]. 中国图象图形学报, 2025, 30(11): 3535-3546.
	Qian Yujie, Ding Dandan. Dynamic point cloud compression method based on inter frame adaptive prediction[J]. Journal of Image and Graphics, 2025, 30(11): 3535-3546.
[20]	Mallick S S, Goel R, Kerbl B, et al. Taming 3DGS: high-quality radiance fields with limited resources[C]//Proceedings of 2024 SIGGRAPH Asia. Tokyo: ACM Press, 2024: 1-11.
[21]	Fan Zhiwen, Wang K, Wen Kairun, et al. LightGaussian: unbounded 3D Gaussian compression with 15x reduction and 200+FPS[J]. Advances in neural information processing systems, 2024, 37: 140138-140158.
[22]	Niedermayr S, Stumpfegger J, Westermann R. Compressed 3D Gaussian splatting for accelerated novel view synthesis[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 10349-10358.
[23]	Navaneet K L, Pourahmadi Meibodi K, Abbasi Koohpayegani S, et al. CompGS: smaller and faster Gaussian splatting with vector quantization[C]//Proceedings of 2024 European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 330-349.
[24]	Lee J C, Rho D, Sun Xiangyu, et al. Compact 3D Gaussian representation for radiance field[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 21719-21728.
[25]	Yan Ke, Wang Hui, Li Zhuxin, et al. A Large-scale 3D Gaussian reconstruction method for optimized adaptive density control in training resource scheduling[J]. Remote Sensing, 2025, 17(23): 3868.
[26]	Guo Xinli, Zhang Weidong, Liu Ruonan, et al. Motiongs: compact Gaussian splatting SLAM by motion filter[C]//Proceedings of 2024 International Conference on Robotics, Control and Automation Engineering. Wuhu: IEEE, 2024: 685-692.
[27]	Morgenstern W, Barthel F, Hilsmann A, et al. Compact 3D scene representation via self-organizing Gaussian grids[C]//Proceedings of 2024 European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 18-34.
[28]	Girish S, Gupta K, Shrivastava A. EAGLES: efficient accelerated 3D Gaussians with lightweight encodings[C]//Proceedings of 2024 European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 54-71.
[29]	Liu Yifei, Zhong Zhihang, Zhan Yifan, et al. MaskGaussian: adaptive 3D Gaussian representation from probabilistic masks[C]//Proceedings of 2025 Computer Vision and Pattern Recognition Conference. Nashville:. IEEE, 2025: 681-690.
[30]	Lu Tao, Yu Mulin, Xu Linning, et al. Scaffold-GS: structured 3D Gaussians for view-adaptive rendering[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 20654-20664.

方法	fr1/desk	fr2/xyz	fr3/Office	Avg
SplaTAM	3.38	1.29	5.14	3.27
Photo-SLAM	1.52	0.34	1.23	1.03
MonoGS-SLAM	1.47	1.54	1.41	1.47
GS-ICP SLAM	2.54	1.78	3.32	2.55
本文方法	1.57	0.43	1.21	1.07

方法	Room0	Room1	Room2	Office0	Office1	Office2	Office3	Office4	Avg
SplaTAM	0.31	0.39	0.29	0.48	0.25	0.28	0.32	0.56	0.36
Photo-SLAM	0.59	0.51	0.32	0.45	0.46	0.48	0.61	0.87	0.54
MonoGS SLAM	0.48	0.34	0.29	0.46	0.49	0.22	0.46	2.43	0.64
GS-ICP SLAM	0.15	0.16	0.11	0.18	0.12	0.17	0.16	0.21	0.16
本文方法	0.14	0.16	0.11	0.18	0.13	0.17	0.16	0.21	0.16

方法	指标	fr1/desk	fr2/xyz	fr3/Office	Avg
SplaTAM	PSNR/dB↑	21.03	23.03	20.63	21.56
	SSIM↑	0.75	0.79	0.73	0.76
	LPIPS↓	0.23	0.20	0.21	0.21
Photo-SLAM	PSNR/dB↑	19.45	21.45	22.63	21.18
	SSIM↑	0.70	0.73	0.76	0.73
	LPIPS↓	0.25	0.16	0.15	0.19
MonoGS-SLAM	PSNR/dB↑	18.66	15.81	19.11	17.86
	SSIM↑	0.70	0.72	0.97	0.80
	LPIPS↓	0.33	0.25	0.32	0.30
GS-ICP SLAM（限制帧率25帧/s）	PSNR/dB↑	17.42	18.43	20.01	18.62
	SSIM↑	0.64	0.68	0.73	0.68
	LPIPS↓	0.37	0.30	0.27	0.31
本文方法	PSNR/dB↑	20.54	22.71	21.95	21.73
	SSIM↑	0.72	0.76	0.74	0.74
	LPIPS↓	0.27	0.19	0.18	0.21

方法	指标	Room0	Room1	Room2	Office0	Office1	Office2	Office3	Office4	Avg
SplaTAM	PSNR/dB↑	32.81	33.55	34.83	38.09	39.02	31.95	29.53	31.55	33.89
	SSIM↑	0.975	0.969	0.982	0.982	0.982	0.966	0.949	0.951	0.970
	LPIPS↓	0.070	0.097	0.074	0.088	0.093	0.098	0.119	0.150	0.099
Photo-SLAM	PSNR/dB↑	33.38	34.10	36.32	38.72	39.31	34.22	34.10	34.82	35.62
	SSIM↑	0.979	0.977	0.985	0.985	0.987	0.962	0.963	0.981	0.977
	LPIPS↓	0.097	0.115	0.101	0.089	0.110	0.152	0.119	0.131	0.114
MonoGS-SLAM	PSNR/dB↑	31.56	32.86	32.59	38.70	41.17	32.36	32.03	32.92	34.27
	SSIM↑	0.968	0.973	0.971	0.986	0.993	0.978	0.970	0.968	0.975
	LPIPS↓	0.094	0.075	0.093	0.050	0.033	0.094	0.110	0.112	0.082
GS-ICP SLAM	PSNR/dB↑	35.43	37.46	38.33	43.00	43.32	36.90	36.94	38.71	38.76
	SSIM↑	0.962	0.969	0.973	0.985	0.984	0.973	0.969	0.972	0.973
	LPIPS↓	0.049	0.047	0.050	0.027	0.031	0.044	0.042	0.046	0.042
本文方法	PSNR/dB↑	36.12	38.10	39.45	43.52	44.13	37.58	37.34	39.19	39.44
	SSIM↑	0.967	0.974	0.978	0.9870	0.987	0.977	0.972	0.975	0.977
	LPIPS↓	0.042	0.040	0.041	0.022	0.021	0.037	0.037	0.040	0.034

方法	指标	InSTaR0	InSTaR1	InSTaR2	Avg
SplaTAM	PSNR/dB↑	19.69	21.54	19.82	20.35
	SSIM↑	0.70	0.72	0.76	0.73
	LPIPS↓	0.37	0.29	0.33	0.33
Photo-SLAM	PSNR/dB↑	23.62	22.47	22.27	22.79
	SSIM↑	0.85	0.81	0.78	0.81
	LPIPS↓	0.18	0.21	0.28	0.22
MonoGS-SLAM	PSNR/dB↑	20.04	19.21	18.23	19.16
	SSIM↑	0.78	0.76	0.72	0.75
	LPIPS↓	0.42	0.25	0.52	0.40
GS-ICP SLAM	PSNR/dB↑	21.06	20.21	19.71	20.33
	SSIM↑	0.89	0.81	0.82	0.84
	LPIPS↓	0.19	0.31	0.39	0.30
本文方法	PSNR↑	24.31	22.31	23.43	23.35
	SSIM↑	0.82	0.79	0.83	0.81
	LPIPS↓	0.17	0.25	0.32	0.25