渐进式学习策略下的无监督三维重建方法

doi:10.12305/j.issn.1001-506X.2026.09.04

系统工程与电子技术 ›› 2026, Vol. 48 ›› Issue (9): 2917-2924.doi: 10.12305/j.issn.1001-506X.2026.09.04

• 电子技术 • 上一篇

渐进式学习策略下的无监督三维重建方法

崔浩浩¹(), 邸彦强¹(), 谢志英¹(), 刘青¹^,²

1. 陆军工程大学石家庄校区，河北石家庄 050003
2. 河北科技大学经济管理学院，河北石家庄 050018

收稿日期:2025-07-27 修回日期:2025-10-12 出版日期:2025-12-08 发布日期:2025-12-08
通讯作者: 邸彦强 E-mail:cuihaosim@aeu.edu.cn;yq.simu@139.com;6135145@qq.com
作者简介:崔浩浩（1987—），男，讲师，博士研究生，主要研究方向为武器系统仿真、人工智能
邸彦强（1973—），男，教授，博士，主要研究方向为武器系统仿真
谢志英（1978—），女，副教授，硕士，主要研究方向为武器系统仿真
刘　青（1982—），男，工程师，博士研究生，主要研究方向为武器系统仿真

Unsupervised three-dimensional reconstruction via progressive learning strategy

Haohao Cui¹(), Yanqiang Di¹(), Zhiying Xie¹(), Qing Liu¹^,²

1. Shijiazhuang Campus，Army Engineering University of PLA，Shijiazhuang 050003，China
2. School of Economics and Management，Hebei University of Science and Technology，Shijiazhuang 050018，China

Received:2025-07-27 Revised:2025-10-12 Online:2025-12-08 Published:2025-12-08
Contact: Yanqiang Di E-mail:cuihaosim@aeu.edu.cn;yq.simu@139.com;6135145@qq.com

摘要/Abstract

摘要：

无监督多视图立体（multi-view stereo, MVS）重建方法无法有效解决背景区域对训练损失造成的干扰，所依赖的光度一致性以及深度平滑性损失在复杂场景中存在失效的情况，导致损失值不可靠的问题。为解决这些问题引入了单目深度估计值遮蔽背景区域，在数据增强手段的基础上，引入由易至难的渐进式学习策略，并利用单目深度值改进了深度平滑损失函数。在三维重建数据集上进行了实验，取得了超越现有无监督MVS方法的实验结果。实验证明了所提出的改进措施能有效提升训练过程的鲁棒性，并有效提升重建质量。

关键词: 多视图立体重建, 无监督多视图立体重建, 单目深度, 渐进式学习策略, 深度平滑损失

Abstract:

Unsupervised multi-view stereo （MVS） reconstruction methods cannot effectively address the interference of background regions on training losses. The relied photometric consistency and depth smoothness losses do not always hold in complex regions, inevitably leading to unreliable loss values. To address these issues, monocular depth estimation values are introduced to mask the background regions. Based on the data augmentation, a progressive learning strategy from easy to difficult is adopted and the depth smoothness loss function is improved by using the monocular depth. Experiments are conducted on the three dimensional reconstruction datasets, achieving results that surpass the existing unsupervised MVS methods. The results prove that the proposed improvement measures can effectively enhance the robustness of the training process and improve the reconstruction quality.

Key words: multi-view stereo reconstruction, unsupervised multi-view stereo (MVS) reconstruction, monocular depth, progressive learning strategy, depth smoothness loss

中图分类号:

TP 391.9

崔浩浩, 邸彦强, 谢志英, 刘青. 渐进式学习策略下的无监督三维重建方法[J]. 系统工程与电子技术, 2026, 48(9): 2917-2924.

Haohao Cui, Yanqiang Di, Zhiying Xie, Qing Liu. Unsupervised three-dimensional reconstruction via progressive learning strategy[J]. Systems Engineering and Electronics, 2026, 48(9): 2917-2924.

图/表 10

图1

图2

图3

表1

图4

表2

表3

表4

图5

图6

参考文献 26

1	樊铭瑞, 申冰可, 牛文龙, 等. 基于深度学习的多视图立体视觉综述[J]. 软件学报, 2025, 36 (4): 1692. doi: 10.13328/j.cnki.jos.007248
2	彭勇, 王爱迪, 王婷婷, 等. 面向生物控制的鲤鱼脑组织及脑电极三维重建[J]. 生物医学工程学杂志, 2020, 37 (5): 885. doi: 10.7507/1001-5515.201911011
3	缪长蔚, 陈铭, 张淑芳, 等. 基于三维重建的虚拟现实类鸟飞行系统[J]. 光电子·激光, 2023, 34 (3): 268.
4	Qu Y F, Huang J Y, Zhang X. Rapid 3D reconstruction for image sequence acquired from UAV camera[J]. Sensors, 2018, 18 (1): 225. doi: 10.3390/s18010225
5	刘鑫达, 李康, 耿国华, 等. 文化遗产保护中虚拟现实技术研究进展[J]. 计算机辅助设计与图形学学报, 2025, 37 (4): 545. doi: 10.3724/SP.J.1089.2024-00509
6	Yao Y, Luo Z X, Li S W, et al. MVSNet: depth inference for unstructured multi-view stereo[C]// European Conference on Computer Vision, 2018: 767.
7	Annaes H, Jensen R R, Vogiatzis G, et al. Large-scale data for multiple-view stereopsis[J]. International Journal of Computer Vision, 2016, 120 (2): 153. doi: 10.1007/s11263-016-0902-9
8	Knapitsch A, Park J, Zhou Q Y, et al. Tanks and temples: benchmarking large-scale scene reconstruction[J]. ACM Transactions on Graphics, 2017, 36 (4): 1.
9	Mahjourian R, Wicke M, Angelova A. Unsupervised learning of depth and ego-motion from monocular video using 3D geometric constraints[C]// IEEE Conference on Computer Vision and Pattern Recognition, 2018: 5667.
10	Khot T, Agrawal S, Tulsiani S, et al. Learning unsupervised multi-view stereopsis via robust photometric consistency[EB/OL]. [2025-05-27]. https://arxiv.org/abs/1905.02706.
11	Xu H B, Zhou Z P, Qiao Y, et al. Self-supervised multi-view stereo via effective co-segmentation and data-augmentation[C]// AAAI Conference on Artificial Intelligence, 2021, 35(4): 3030.
12	Yao Y, Luo Z X, Li S W, et al. Recurrent MVSNet for high-resolution multi-view stereo depth inference[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 5525.
13	Wang F, Galliani S, Vogel C, et al. PatchmatchNet: learned multi-view patchmatch stereo[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 14194.
14	Gu X D, Fan Z W, Zhu S Y, et al. Cascade cost volume for high-resolution multi-view stereo and stereo matching[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 2495.
15	Zhang Z, Peng R, Hu Y X, et al. GeoMVSNet: learning multi-view stereo with geometry perception[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023: 21508.
16	Chang D, Božič A, Zhang T, et al. RC-MVSNet: unsupervised multi-view stereo with neural rendering[C]// European Conference on Computer Vision, 2022: 665.
17	Wang Z, Bovik A C, Sheikh H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004, 13 (4): 600. doi: 10.1109/TIP.2003.819861
18	Huang B, Yi H, Huang C, et al. M3VSNet: unsupervised multi-metric multi-view stereo network[C]// IEEE International Conference on Image Processing, 2021: 3163.
19	Xiong K Q, Peng R, Zhang Z, et al. CL-MVSNet: unsupervised multi-view stereo with dual-level contrastive learning[C]// IEEE/CVF International Conference on Computer Vision, 2023: 3769.
20	罗锦锋, 袁冬莉, 张蓝, 等. 基于多视图传播的无监督三维重建方法[J]. 西北工业大学学报, 2024, 42 (1): 129. doi: 10.1051/jnwpu/20244210129
21	Yang L H, Kang B Y, Huang Z L, et al. Depth anything V2[J]. Advances in Neural Information Processing Systems, 2024, 37, 21875.
22	Cheng S, Xu Z X, Zhu S L, et al. Deep stereo using adaptive thin volume representation with uncertainty awareness[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 2524.
23	Yang J Y, Alvarez J, Liu M M. Self-supervised learning of depth inference for multi-view stereo[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 7526.
24	Xu H B, Zhou Z P, Wang Y L, et al. Digging into uncertainty in self-supervised multi-view stereo[C]// IEEE/CVF International Conference on Computer Vision, 2021: 6078.
25	Ding Y K, Zhu Q T, Liu X Y, et al. KD-MVS: knowledge distillation based self-supervised learning for MVS[EB/OL]. [2025-05-27]. https://arxiv.org/abs/2207.10425.
26	Dai Y C, Zhu Z D, Rao Z B, et al. MVS2: deep unsupervised multi-view stereo with multi-view symmetry[C]// International Conference on 3D Vision, 2019.

类别	方法	Acc	Comp	Overall
有监督	MVSNet	0.396	0.527	0.462
	CasMVSNet	0.325	0.385	0.355
	PatchmatchNet	0.427	0.277	0.352
	UCS-Net	0.338	0.349	0.344
	GeoMVSNet	0.331	0.259	0.295
自监督	Self-sup CVP	0.308	0.418	0.363
	U-MVSNet	0.354	0.353	0.353
	KD-MVS	0.359	0.295	0.327
无监督	Unsup-MVSNet	0.881	1.073	0.977
	JDACS	0.398	0.318	0.358
	RC-MVSNet	0.396	0.295	0.345
	CL-MVSNet	0.375	0.283	0.329
	本方法	0.372	0.279	0.325

类别	方法	均值	Family	Francis	Horse	Lighthouse	M60	Panther	Playground	Train
有监督	MVSNet	43.48	55.09	28.55	25.07	50.79	53.96	50.86	47.90	34.69
	PatchmatchNet	53.15	66.99	52.64	43.24	54.87	52.87	49.54	54.21	50.81
	UCS-Net	54.83	76.09	53.16	43.03	54.00	55.60	51.49	57.38	47.89
	CasMVSNet	56.42	76.36	58.45	46.20	55.53	56.11	54.02	58.17	46.56
	GeoMVSNet	65.89	81.64	67.53	55.78	68.02	65.49	67.19	63.27	58.22
自监督	Self-sup CVP	46.71	64.95	38.79	24.98	49.73	52.57	51.53	50.66	40.45
	U-MVSNet	57.15	76.49	60.04	49.20	55.52	55.33	51.22	56.77	52.63
	KD-MVS	64.14	80.42	67.42	54.02	64.52	64.18	61.60	62.37	58.59
无监督	MVS2	37.21	47.74	21.55	19.50	44.54	44.86	46.32	43.38	29.72
	JDACS	45.48	66.62	38.25	36.11	46.12	46.66	42.25	47.69	37.16
	RC-MVSNet	55.04	75.26	53.50	45.52	53.49	54.85	52.30	56.06	49.37
	CL-MVSNet	59.39	76.35	62.37	49.93	60.02	57.44	59.97	56.74	52.28
	本方法	59.65	78.36	63.38	51.35	59.43	59.00	53.73	58.02	53.95

类别	方法	均值	Auditorium	Ballroom	Courtroom	Museum	Palace	Temple
有监督	MVSNet	—	—	—	—	—	—	—
	PatchmatchNet	32.31	23.69	37.73	30.04	41.80	28.31	32.29
	UCS-Net	—	—	—	—	—	—	—
	CasMVSNet	31.12	19.81	38.46	29.10	43.87	27.36	28.11
	GeoMVSNet	41.52	30.23	46.53	39.98	53.05	35.98	43.34
自监督	Self-sup CVP	—	—	—	—	—	—	—
	U-MVSNet	—	—	—	—	—	—	—
	KD-MVS	37.96	27.24	44.10	35.47	49.16	34.68	37.11
无监督	MVS2	—	—	—	—	—	—	—
	JDACS	—	—	—	—	—	—	—
	RC-MVSNet	30.82	21.72	37.22	28.62	37.37	27.88	32.09
	CL-MVSNet	37.03	28.07	43.55	37.47	50.86	31.45	30.78
	本方法	38.57	25.35	44.40	37.08	50.58	33.17	40.84

${L_{CS}}$	BR	PL	${L_{DS}}$	Acc	Comp	Overall
√	—	—	—	0.422	0.334	0.378
√	√	—	—	0.391	0.316	0.354
√	√	√	—	0.370	0.303	0.336
—	√	√	√	0.372	0.279	0.325

渐进式学习策略下的无监督三维重建方法

Unsupervised three-dimensional reconstruction via progressive learning strategy

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 26

相关文章 1

编辑推荐

Metrics

本文评价