基于共视引导的大规模场景运动恢复结构

doi:10.11996/JG.j.2095-302X.2026040695

摘要/Abstract

摘要：

随着混合现实与自动驾驶等空间智能应用的发展，对大规模场景下高精度三维重建技术的需求日益增长。在大规模场景中，传统的运动恢复结构方法计算效率低，误差累积严重。针对现有的分块重建方法无法规避累积误差带来的精度损失的问题，提出一种基于共视引导的大规模场景运动恢复结构方法。在全局层级，利用贪心策略筛选关键帧，通过共视图自适应加权融合GPS与视觉信息，快速生成全局一致的稀疏场景骨架。在局部层级，利用共视关系将场景自适应划分为若干视觉高内聚的图像社区从而进行局部重建。在模型融合阶段，设计基于双向一致性检验的模型融合策略，剔除错误相对变换，并通过多源变换联合优化将所有局部模型在全局骨架约束下对齐。实验结果表明，该方法能够显著提升大场景重建的效率和鲁棒性，有效抑制累计误差。

关键词: 三维重建, 运动恢复结构, 分层重建, GPS融合, 模型融合

Abstract:

With the advancement of spatial intelligence applications such as mixed reality and autonomous driving, demand for high-precision 3D reconstruction technologies in large-scale scenes has continued to increase. In such scenarios, traditional Structure from Motion (SfM) methods are computationally inefficient and prone to severe error accumulation, while existing divide-and-conquer reconstruction strategies fail to mitigate the accuracy loss caused by cumulative errors. To address these issues, a covisibility-guided hierarchical partitioning SfM method for large-scale scenes was proposed. At the global level, keyframes were selected using a greedy strategy, and GPS and visual information were fused via an adaptive covisibility weighting mechanism to rapidly generate a globally consistent sparse scene skeleton. At the local level, the scene was adaptively partitioned into several visually high-cohesion image communities based on covisibility relationships for local reconstruction. In the model fusion stage, a fusion strategy based on bidirectional consistency checking was designed to eliminate erroneous relative transformations, and all local models were aligned under global-skeleton constraints through joint optimization of multi-source transformations. Experimental results indicated that the proposed method significantly improved the efficiency and robustness of large-scale scene reconstruction and effectively suppressed cumulative errors.

Key words: 3D reconstruction, structure from motion, hierarchical reconstruction, global positioning system fusion, model fusion

中图分类号:

TP391

欧阳泽洪, 沈旭昆, 任曦, 胡勇, 黄勇. 基于共视引导的大规模场景运动恢复结构[J]. 图学学报, 2026, 47(4): 695-703.

OUYANG Zehong, SHEN Xukun, REN Xi, HU Yong, HUANG Yong. Covisibility-based large-scale structure from motion[J]. Journal of Graphics, 2026, 47(4): 695-703.

图/表 14

参考文献 22

[1]	ZHOU L L, WU G X, ZUO Y B, et al. A comprehensive review of vision-based 3D reconstruction methods[J]. Sensors, 2024, 24(7): 2314. DOI URL
[2]	GUO H Y, ZHU H, PENG S D, et al. Multi-view reconstruction via SfM-guided monocular depth estimation[C]// 2025 IEEE/ CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2025: 5272-5282.
[3]	葛海明, 张维, 王小龙, 等. 基于SfM的城市电缆隧道三维重建方法优化研究[J]. 图学学报, 2023, 44(3): 540-550. DOI
	GE H M, ZHANG W, WANG X L, et al. Optimization of 3D reconstruction method for urban cable tunnel based on SfM method[J]. Journal of Graphics, 2023, 44(3): 540-550 (in Chinese).
[4]	XIAO Y X, XUE N, WU T F, et al. Level-S²fM: structure from motion on neural level set of implicit surfaces[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2023: 17205-17214.
[5]	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.
[6]	KERBL B, KOPANAS G, LEIMKUEHLER T, et al. 3D Gaussian splatting for real-time radiance field rendering[J]. ACM Transactions on Graphics, 2023, 42(4): 139.
[7]	黄志勇, 佘雅丽, 华喜锋, 等. DCSplat: 一种深度约束的稀疏视角三维重建方法[J]. 图学学报, 2025, 46(3): 510-519. DOI
	HUANG Z Y, SHE Y L, HUA X F, et al. DCSplat: Gaussian splatting with depth information constraints under sparse viewpoints[J]. Journal of Graphics, 2025, 46(3): 510-519 (in Chinese). DOI
[8]	高翔, 李梦晗, 申抒含. 大规模场景运动恢复结构研究综述[J]. 计算机辅助设计与图形学学报, 2024, 36(7): 969-994.
	GAO X, LI M H, SHEN S H. Large-scale structure from motion: a survey[J]. Journal of Computer-Aided Design & Computer Graphics, 2024, 36(7): 969-994. (in Chinese)
[9]	AGARWAL S, SNAVELY N, SIMON I, et al. Building Rome in a day[C]// The 12th International Conference on Computer Vision. New York: IEEE Press, 2009: 72-79.
[10]	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.
[11]	WILSON K, SNAVELY N. Robust global translations with 1DSFM[C]// The 13th European Conference on Computer Vision. Cham: Springer, 2014: 61-75.
[12]	CUI H N, GAO X, SHEN S H, et al. HSfM: hybrid structure-from-motion[C]// IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2017: 2393-2402.
[13]	GALVEZ-LÓPEZ D, TARDOS J D. Bags of binary words for fast place recognition in image sequences[J]. IEEE Transactions on Robotics, 2012, 28(5): 1188-1197. DOI URL
[14]	BHOWMICK B, PATRA S, CHATTERJEE A, et al. Divide and conquer: a hierarchical approach to large-scale structure- from-motion[J]. Computer Vision and Image Understanding, 2017, 157: 190-205. DOI URL
[15]	CHEN Y, SHEN S H, CHEN Y S, et al. Graph-based parallel large scale structure from motion[J]. Pattern Recognition, 2020, 107: 107537. DOI URL
[16]	BLONDEL V D, GUILLAUME J L, LAMBIOTTE R, et al. Fast unfolding of communities in large networks[J]. Journal of Statistical Mechanics: Theory and Experiment, 2008, 2008: P10008. DOI
[17]	HUANG W, JIANG S, JIANG W S. Camera self-calibration with GNSS constrained bundle adjustment for weakly structured long corridor UAV images[J]. Remote Sensing, 2021, 13(21): 4222. DOI URL
[18]	VENATOR M, BRUNS E, MAIER A. Robust camera pose estimation for unordered road scene images in varying viewing conditions[J]. IEEE Transactions on Intelligent Vehicles, 2020, 5(1): 165-174. DOI URL
[19]	GEIGER A, LENZ P, STILLER C, et al. Vision meets robotics: the KITTI dataset[J]. The International Journal of Robotics Research, 2013, 32(11): 1231-1237. DOI URL
[20]	CAI R J, TUNG J, WANG Q Q, et al. Doppelgangers: learning to disambiguate images of similar structures[C]// 2013 IEEE/ CVF International Conference on Computer Vision. New York: IEEE Press, 2023: 34-44.
[21]	WANG L, GE L L, LUO S, et al. TC-SfM: robust track- community-based structure-from-motion[J]. IEEE Transactions on Image Processing, 2024, 33: 1534-1548. DOI URL
[22]	GRISETTI G, KÜMMERLE R, STACHNISS C, et al. A tutorial on graph-based SLAM[J]. IEEE Intelligent Transportation Systems Magazine, 2010, 2(4): 31-43.

序列号	图像数量	姿态角误差的中值误差与平均误差/m
序列号	图像数量	CM		GS		DG				TS				Ours
00	9 082	5.060	5.020	2.450	16.780	0.980			1.430	0.980	1.010			0.840	1.020
01	2 202	0.560	0.590	0.570	0.890	0.600			0.940	0.700	0.950			0.510	0.520
02	9 322	0.680	0.830	1.040	1.370	1.350			1.470	1.150	1.330			0.560	0.640
03	1 602	0.170	0.200	0.360	0.420	0.200			0.230	0.370	0.430			0.110	0.120
04	542	1.740	1.730	47.990	50.430	2.620			2.620	3.070	3.070			0.850	0.930
05	5 522	2.650	2.550	1.690	2.220	1.230			2.000	0.460	0.550			0.400	0.430
06	2 202	1.710	2.800	1.960	2.590	1.610			2.600	0.590	1.610			0.510	0.650
07	2 202	5.820	5.730	2.140	2.130	2.160			2.070	0.930	1.030			0.810	0.870
08	8 142	1.730	1.700	5.770	6.110	2.120			2.050	2.430	2.740			1.020	1.330
09	3 182	4.020	4.050	2.040	2.140	1.900			1.700	0.450	0.540			0.500	0.530
10	2 402	1.260	1.300	0.910	1.460	0.850			1.430	0.940	1.060			0.810	1.010
序列号	图像数量	平移误差中的中值误差与平均误差/m
序列号	图像数量	CM		GS			DG			TS			Ours
00	9 082	32.440	36.170	18.780	31.800		3.89	4.53		4.180		5.010	1.140		1.090
01	2 202	14.830	16.810	1.570	1.880		1.690	2.080		1.890		2.010	1.430		1.550
02	9 322	71.870	78.340	11.280	11.710		59.750	66.740		8.620		11.680	1.230		1.340
03	1 602	0.840	0.980	0.320	0.390		0.540	0.710		0.390		0.460	0.410		0.440
04	542	0.140	0.160	29.720	31.100		0.150	0.180		0.160		0.190	0.160		0.180
05	5 522	10.450	12.520	5.280	6.450		3.260	4.320		1.480		1.670	1.110		1.210
06	2 202	11.600	17.060	3.140	4.650		12.410	17.970		1.980		2.800	0.890		1.020
07	2 202	14.410	16.250	2.970	4.040		3.830	4.090		2.440		2.480	0.860		0.970
08	8 142	43.980	61.600	224.110	234.190		47.030	67.460		8.420		10.750	1.330		1.650
09	3 182	31.840	37.140	6.320	6.830		6.170	8.440		2.890		3.600	0.920		1.030
10	2 402	7.180	8.540	1.680	2.240		1.910	2.510		1.800		2.190	0.960		1.230

序列号	图像数量	姿态角误差的中值误差与平均误差/m
序列号	图像数量	CM		GS		DG				TS				Ours
00	9 082	5.060	5.020	2.450	16.780	0.980			1.430	0.980	1.010			0.840	1.020
01	2 202	0.560	0.590	0.570	0.890	0.600			0.940	0.700	0.950			0.510	0.520
02	9 322	0.680	0.830	1.040	1.370	1.350			1.470	1.150	1.330			0.560	0.640
03	1 602	0.170	0.200	0.360	0.420	0.200			0.230	0.370	0.430			0.110	0.120
04	542	1.740	1.730	47.990	50.430	2.620			2.620	3.070	3.070			0.850	0.930
05	5 522	2.650	2.550	1.690	2.220	1.230			2.000	0.460	0.550			0.400	0.430
06	2 202	1.710	2.800	1.960	2.590	1.610			2.600	0.590	1.610			0.510	0.650
07	2 202	5.820	5.730	2.140	2.130	2.160			2.070	0.930	1.030			0.810	0.870
08	8 142	1.730	1.700	5.770	6.110	2.120			2.050	2.430	2.740			1.020	1.330
09	3 182	4.020	4.050	2.040	2.140	1.900			1.700	0.450	0.540			0.500	0.530
10	2 402	1.260	1.300	0.910	1.460	0.850			1.430	0.940	1.060			0.810	1.010
序列号	图像数量	平移误差中的中值误差与平均误差/m
序列号	图像数量	CM		GS			DG			TS			Ours
00	9 082	32.440	36.170	18.780	31.800		3.89	4.53		4.180		5.010	1.140		1.090
01	2 202	14.830	16.810	1.570	1.880		1.690	2.080		1.890		2.010	1.430		1.550
02	9 322	71.870	78.340	11.280	11.710		59.750	66.740		8.620		11.680	1.230		1.340
03	1 602	0.840	0.980	0.320	0.390		0.540	0.710		0.390		0.460	0.410		0.440
04	542	0.140	0.160	29.720	31.100		0.150	0.180		0.160		0.190	0.160		0.180
05	5 522	10.450	12.520	5.280	6.450		3.260	4.320		1.480		1.670	1.110		1.210
06	2 202	11.600	17.060	3.140	4.650		12.410	17.970		1.980		2.800	0.890		1.020
07	2 202	14.410	16.250	2.970	4.040		3.830	4.090		2.440		2.480	0.860		0.970
08	8 142	43.980	61.600	224.110	234.190		47.030	67.460		8.420		10.750	1.330		1.650
09	3 182	31.840	37.140	6.320	6.830		6.170	8.440		2.890		3.600	0.920		1.030
10	2 402	7.180	8.540	1.680	2.240		1.910	2.510		1.800		2.190	0.960		1.230

图像数量	总耗时	全局级SfM	社区划分	局部级SfM	融合
532	18.51	7.68	0.80	5.30	5.23
1 346	226.68	91.93	4.38	60.66	65.79
2 012	181.66	77.07	3.71	48.43	53.25
3 116	237.63	103.12	4.62	64.42	64.43
3 753	404.64	166.62	8.11	112.53	110.93
4 435	361.85	153.68	6.98	99.21	99.96
4 861	498.09	208.28	10.25	135.28	139.66
5 230	413.62	175.33	7.90	117.06	112.00
5 869	517.95	208.08	10.82	151.78	149.50
6 124	636.63	262.16	12.22	181.54	177.19

图像数量	总耗时	全局级SfM	社区划分	局部级SfM	融合
532	18.51	7.68	0.80	5.30	5.23
1 346	226.68	91.93	4.38	60.66	65.79
2 012	181.66	77.07	3.71	48.43	53.25
3 116	237.63	103.12	4.62	64.42	64.43
3 753	404.64	166.62	8.11	112.53	110.93
4 435	361.85	153.68	6.98	99.21	99.96
4 861	498.09	208.28	10.25	135.28	139.66
5 230	413.62	175.33	7.90	117.06	112.00
5 869	517.95	208.08	10.82	151.78	149.50
6 124	636.63	262.16	12.22	181.54	177.19

序列号	Visual-only				Fixed-weight				Ours
序列号	e_trans/m	e_rot/°	Cover/%	T/min	e_trans/m	e_rot/°	Cover/%	T/min	e_trans/m	e_rot/°	Cover/%	T/min
01	3.15	0.85	98	6.65	0.64	0.88	99	8.66	0.42	0.51	100	7.68
02	5.82	0.91	95	90.85	0.61	0.92	99	96.90	0.55	0.51	100	91.93
03	—	—	0	-	0.77	0.95	96	79.01	0.61	0.54	99	77.07
04	6.55	0.94	94	101.05	0.76	0.96	97	118.10	0.65	0.56	100	103.12
05	8.91	0.98	91	163.51	0.88	0.97	96	160.58	0.73	0.60	100	166.62
06	8.13	0.96	92	148.55	0.94	0.97	96	158.62	0.83	0.71	100	153.68
07	—	—	0	-	0.85	0.81	96	211.15	0.78	0.63	100	208.28
08	9.32	0.99	90	178.00	0.96	0.85	93	177.29	0.76	0.66	98	175.33
09	—	—	0	-	0.93	0.84	93	207.98	0.77	0.66	97	208.08
10	—	—	0	-	0.94	0.86	93	262.10	0.79	0.67	99	262.16