融合层次化双流注意力的单目深度估计方法

doi:10.11996/JG.j.2095-302X.2026030553

图学学报 ›› 2026, Vol. 47 ›› Issue (3): 553-563.DOI: 10.11996/JG.j.2095-302X.2026030553

• 图像处理与计算机视觉 • 上一篇下一篇

融合层次化双流注意力的单目深度估计方法

吴文欢¹^,²^,³(), 王文舒¹, 王舒鳌¹

¹ 湖北汽车工业学院，智能网联汽车学院，湖北十堰 442002
² 湖北汽车工业学院，汽车动力传动与电子控制湖北省重点实验室，湖北十堰 442002
³ 湖北汽车工业学院，空地群智协同技术与应用十堰市重点实验室，湖北十堰 442002

收稿日期:2025-07-04 接受日期:2026-02-02 出版日期:2026-06-30 发布日期:2026-06-30
通讯作者:吴文欢，E-mail：wuwenhuan5@163.com
基金资助:
湖北省自然科学基金创新发展联合基金项目(2025AFD239);汽车动力传动与电子控制湖北省重点实验室开放基金项目(ZDK12025A02)

Monocular depth estimation method with hierarchical dual-stream attention

WU Wenhuan¹^,²^,³(), WANG Wenshu¹, WANG Shuao¹

¹ School of Intelligent Connected Vehicle, Hubei University of Automotive Technology, Shiyan Hubei 442002, China
² Key Laboratory of Automotive Power Train and Electronics, Hubei University of Automotive Technology, Shiyan Hubei 442002, China
³ Shiyan Key Laboratory of Air-Ground Crowd Cooperation Technology and Application, Hubei University of Automotive Technology, Shiyan Hubei 442002, China

Received:2025-07-04 Accepted:2026-02-02 Published:2026-06-30 Online:2026-06-30
Contact: WU Wenhuan，E-mail：wuwenhuan5@163.com
Supported by:
Innovation and Development Joint Fund Project of Natural Science Foundation of Hubei Province(2025AFD239);Open Fund Project of Key Laboratory of Automotive Power Train and Electronics (Hubei University of Automotive Technology)(ZDK12025A02)

摘要/Abstract

摘要：

单目深度估计因其在自动驾驶、三维重建等领域具有广泛地应用前景而持续受到关注。然而，现有方法在多尺度特征融合与深度建模方面仍存在不足，难以同时兼顾局部几何细节刻画与全局结构一致性，对不同场景下深度尺度分布的自适应能力有限。针对上述问题，构建了一种融合层次化双流注意力与自适应深度离散建模的单目深度估计框架。编码阶段采用 Swin Transformer 构建金字塔式多尺度特征表示，以增强对局部与全局信息的联合建模能力；解码阶段设计层次化双流注意力融合网络，在逐级重建过程中并行建模局部细节感知与全局上下文语义，并通过动态权重调制与交叉注意力机制实现特征的自适应融合。同时，引入深度恢复模块，将深度估计建模为分类与回归相结合的任务，通过预测离散深度分布并自适应学习分箱中心，以概率加权方式生成连续深度结果，在保证深度预测连续性的同时，有效保持场景中深度关系的一致性。实验结果表明，该方法在 KITTI 数据集上的AbsRel 为 0.048、SqRel 为 0.147、log10 为 0.020、δ? 为 0.980，并在 NYU Depth V2 数据集上展现出良好的泛化能力，验证了该方法在复杂场景和多尺度深度分布条件下的有效性与鲁棒性。

关键词: 深度学习, 单目深度估计, 金字塔结构, 注意力机制, 特征融合

Abstract:

Monocular depth estimation has attracted sustained attention due to its broad application prospects in autonomous driving, 3D reconstruction, and related fields. However, existing methods still exhibit deficiencies in multi-scale feature fusion and depth modeling, which makes it difficult to simultaneously capture fine-grained local geometric details and maintain global structural consistency, and have limited adaptability to depth scale distributions across different scenes. To address these issues, a monocular depth estimation framework integrating hierarchical dual-stream attention and adaptive depth discretization was developed. In the encoding stage, a Swin Transformer was employed to construct pyramid multi-scale feature representations, enhancing the joint modeling of local and global information. In the decoding stage, a hierarchical dual-stream attention fusion network was designed to model local detail perception and global contextual semantics in parallel during progressive reconstruction, in which adaptive feature fusion was achieved through dynamic weight modulation and cross-attention mechanisms. Meanwhile, a depth recovery module was introduced to formulate depth estimation as a joint classification-regression task. By predicting a discrete depth distribution and adaptively learning the bin centers, continuous depth values were generated via probability-weighted aggregation, which ensured the continuity of depth prediction while effectively maintaining the consistency of depth relationships within the scene. Experimental results demonstrated that the proposed method achieved state-of-the-art performance on the KITTI dataset, with an AbsRel of 0.048, a SqRel of 0.147, a log10 error of 0.020, and a δ? accuracy of 0.980, and exhibited favorable generalization capability on the NYU Depth V2 dataset, validating its effectiveness and robustness under complex scenes and multi-scale depth distributions.

Key words: deep learning, monocular depth estimation, pyramid structure, attention mechanism, feature fusion

中图分类号:

TP391

吴文欢, 王文舒, 王舒鳌. 融合层次化双流注意力的单目深度估计方法[J]. 图学学报, 2026, 47(3): 553-563.

WU Wenhuan, WANG Wenshu, WANG Shuao. Monocular depth estimation method with hierarchical dual-stream attention[J]. Journal of Graphics, 2026, 47(3): 553-563.

图/表 8

参考文献 37

[1]	LEE Y, JEON J, KO Y, et al. Task-driven deep image enhancement network for autonomous driving in bad weather[C]//2021 IEEE International Conference on Robotics and Automation. New York: IEEE Press, 2021: 13746-13753.
[2]	CHO J, RAHIMPOUR S, CUTLER A, et al. Enhancing reality: a systematic review of augmented reality in neuronavigation and education[J]. World Neurosurgery, 2020, 139: 186-195. DOI PMID
[3]	BROOKS S P. Markov chain Monte Carlo method and its application[J]. Journal of the Royal Statistical Society. Series D (The Statistician), 1998, 47(1): 69-100.
[4]	EIGEN D, PUHRSCH C, FERGUS R. Depth map prediction from a single image using a multi-scale deep network[C]//The 28th International Conference on Neural Information Processing Systems. Cambridge: MIT Press, 2014: 2366-2374.
[5]	LAINA I, RUPPRECHT C, BELAGIANNIS V, et al. Deeper depth prediction with fully convolutional residual networks[C]//2016 Fourth international conference on 3D vision. New York: IEEE Press, 2016: 239-248.
[6]	ARAMPATZAKIS V, PAVLIDIS G, MITIANOUDIS N, et al. Monocular depth estimation: a thorough review[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(4): 2396-2414. DOI URL
[7]	RANFTL R, BOCHKOVSKIY A, KOLTUN V. Vision transformers for dense prediction[C]//2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 12159-12168.
[8]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words:transformers for image recognition at scale[EB/OL]. (2020-10-22) [2025-01-29]. https://arxiv.org/pdf/2010.11929.pdf.
[9]	LI Z Y, CHEN Z H, LIU X M, et al. Depthformer: exploiting long-range correlation and local information for accurate monocular depth estimation[J]. Machine Intelligence Research, 2023, 20(6): 837-854. DOI
[10]	RANFTL R, LASINGER K, HAFNER D, et al. Towards robust monocular depth estimation: mixing datasets for zero-shot cross-dataset transfer[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(3): 1623-1637. DOI URL
[11]	BHAT S F, ALHASHIM I, WONKA P. AdaBins: depth estimation using adaptive bins[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 4008-4017.
[12]	AGARWAL A, ARORA C. Attention attention everywhere: monocular depth prediction with skip attention[C]//2023 IEEE/CVF Winter Conference on Applications of Computer Vision. New York: IEEE Press, 2023: 5850-5859.
[13]	YANG L H, KANG B Y, HUANG Z L, et al. Depth anything: unleashing the power of large-scale unlabeled data[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 10371-10381.
[14]	YANG L H, KANG B Y, HUANG Z L, et al. Depth anything V2[C]//The 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 688.
[15]	CHEN S L, GUO H K, ZHU S N, et al. Video depth anything: consistent depth estimation for super-long videos[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2025: 22831-22840.
[16]	ZHANG N, NEX F, VOSSELMAN G, et al. Lite-mono: a lightweight CNN and transformer architecture for self-supervised monocular depth estimation[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2023: 18537-18546.
[17]	LIU Z, LIN Y T, CAO Y, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]//2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 9992-10002.
[18]	OUYANG D L, HE S, ZHANG G Z, et al. Efficient multi-scale attention module with cross-spatial learning[C]//IEEE International Conference on Acoustics, Speech and Signal Processing. New York: IEEE Press, 2023: 1-5.
[19]	DONG X Y, BAO J N, CHEN D D, et al. CSWin transformer: a general vision transformer backbone with cross-shaped windows[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 12114-12124.
[20]	EIGEN D, FERGUS R. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture[C]//2015 IEEE International Conference on Computer Vision. New York: IEEE Press, 2015: 2650-2658.
[21]	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
[22]	SILBERMAN N, HOIEM D, KOHLI P, et al. Indoor segmentation and support inference from rgbd images[C]//The 12th European Conference on Computer Vision. Cham: Springer, 2012: 746-760.
[23]	PASZKE A, GROSS S, MASSA F, et al. PyTorch: an imperative style, high-performance deep learning library[EB/OL]. (2019-12-03) [2025-01-29]. https://arxiv.org/pdf/1912.01703.pdf.
[24]	KINGMA D P, BA J L. Adam: a method for stochastic optimization[EB/OL]. (2014-12-22) [2025-01-29]. https://arxiv.org/pdf/1412.6980.pdf.
[25]	FU H, GONG M M, WANG C H, et al. Deep ordinal regression network for monocular depth estimation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 2002-2011.
[26]	LEE J H, HAN M K, KO D W, et al. From big to small: multi-scale local planar guidance for monocular depth estimation[EB/OL]. (2019-07-25) [2025-01-29]. https://arxiv.org/pdf/1907.10326.pdf.
[27]	YANG G L, TANG H, DING M L, et al. Transformer-based attention networks for continuous pixel-wise prediction[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 16249-16259.
[28]	SHAO S W, LI R, PEI Z C, et al. NENet: monocular depth estimation via neural ensembles[EB/OL]. (2021-11-16) [2025-01-29]. https://arxiv.org/abs/2111.08313v1.
[29]	SHAO S W, LI R, PEI Z C, et al. Towards comprehensive monocular depth estimation: multiple heads are better than one[J]. IEEE Transactions on Multimedia, 2023, 25: 7660-7671. DOI URL
[30]	PATIL V, SAKARIDIS C, LINIGER A, et al. P3Depth: monocular depth estimation with a piecewise planarity prior[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 1600-1611.
[31]	YUAN W H, GU X D, DAI Z Z, et al. Neural window fully-connected crfs for monocular depth estimation[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 3906-3915.
[32]	BAE J, MOON S, IM S. Deep digging into the generalization of self-supervised monocular depth estimation[C]//The 37th AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2023: 187-196.
[33]	SHIM D, KIM H J. SwinDepth: unsupervised depth estimation using monocular sequences via swin transformer and densely cascaded network[C]//2023 IEEE International Conference on Robotics and Automation. New York: IEEE Press, 2023: 4983-4990.
[34]	WANG B Y, WANG S, YE D, et al. Deep neighbor layer aggregation for lightweight self-supervised monocular depth estimation[C]//ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing. New York: IEEE Press, 2024: 4405-4409.
[35]	HUYNH L, NGUYEN-HA P, MATAS J, et al. Guiding monocular depth estimation using depth-attention volume[C]// The 16th European Conference on Computer Vision. Cham: Springer, 2020: 581-597.
[36]	YIN W, LIU Y F, SHEN C H, et al. Enforcing geometric constraints of virtual normal for depth prediction[C]//2019 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 5683-5692.
[37]	LV C, HAN C G, LANG J, et al. GDM-depth: leveraging global dependency modelling for self-supervised indoor depth estimation[J]. Image and Vision Computing, 2024, 149: 105160. DOI URL

Method	SqRel↓	AbsRel↓	RMSE↓	δ<1.25↑	δ<1.25²↑	δ<1.25³↑
Eigen^[20]	1.548	0.203	6.307	0.769	0.950	0.988
DORN^[25]	0.307	0.072	2.727	0.932	0.984	0.994
BTS^[26]	0.245	0.059	2.756	0.956	0.993	0.998
Adabins^[11]	0.190	0.058	2.360	0.964	0.995	0.999
TransDepth^[27]	0.252	0.064	2.755	0.956	0.994	0.998
NENet^[28]	0.213	0.059	2.543	0.961	0.995	0.999
TEDepth^[29]	0.174	0.056	2.223	0.968	0.996	0.999
P3Depth^[30]	0.270	0.071	2.842	0.953	0.993	0.998
NeWCRFs^[31]	0.155	0.052	2.129	0.974	0.997	0.999
MonoFormer^[32]	0.846	0.104	4.580	0.891	0.962	0.982
SwinDepth^[33]	0.739	0.106	4.510	0.890	0.964	0.984
DepthFormer^[9]	0.158	0.052	2.143	0.975	0.997	0.999
DNA-Depth^[34]	0.682	0.097	4.357	0.902	0.968	0.984
Ours	0.147	0.048	2.094	0.980	0.997	0.999

Method	SqRel↓	AbsRel↓	RMSE↓	δ<1.25↑	δ<1.25²↑	δ<1.25³↑
Eigen^[20]	1.548	0.203	6.307	0.769	0.950	0.988
DORN^[25]	0.307	0.072	2.727	0.932	0.984	0.994
BTS^[26]	0.245	0.059	2.756	0.956	0.993	0.998
Adabins^[11]	0.190	0.058	2.360	0.964	0.995	0.999
TransDepth^[27]	0.252	0.064	2.755	0.956	0.994	0.998
NENet^[28]	0.213	0.059	2.543	0.961	0.995	0.999
TEDepth^[29]	0.174	0.056	2.223	0.968	0.996	0.999
P3Depth^[30]	0.270	0.071	2.842	0.953	0.993	0.998
NeWCRFs^[31]	0.155	0.052	2.129	0.974	0.997	0.999
MonoFormer^[32]	0.846	0.104	4.580	0.891	0.962	0.982
SwinDepth^[33]	0.739	0.106	4.510	0.890	0.964	0.984
DepthFormer^[9]	0.158	0.052	2.143	0.975	0.997	0.999
DNA-Depth^[34]	0.682	0.097	4.357	0.902	0.968	0.984
Ours	0.147	0.048	2.094	0.980	0.997	0.999

Method	AbsRel↓	RMSE↓	log10↓	δ<1.25↑	δ<1.25²↑	δ<1.25³↑
Eigen^[20]	0.158	0.641	-	0.769	0.950	0.988
DORN^[25]	0.115	0.509	0.051	0.828	0.965	0.992
BTS^[26]	0.110	0.392	0.047	0.885	0.978	0.994
TransDepth^[27]	0.106	0.365	0.045	0.900	0.983	0.996
Adabins^[11]	0.103	0.364	0.044	0.903	0.984	0.997
DAV^[35]	0.108	0.412	-	0.882	0.980	0.996
VNL^[36]	0.108	0.416	0.048	0.875	0.976	0.994
TEDepth^[29]	0.100	0.349	0.043	0.907	0.987	0.998
P3Depth^[30]	0.104	0.356	0.043	0.898	0.981	0.996
NENet^[28]	0.100	0.349	0.043	0.907	0.987	0.998
NeWCRFs^[31]	0.095	0.334	0.041	0.922	0.992	0.998
DepthFormer^[9]	0.096	0.339	0.041	0.921	0.989	0.998
PixelFormer^[12]	0.090	0.322	0.039	0.929	0.991	0.998
GDM-Depth^[37]	0.113	0.439	0.049	0.872	0.972	0.993
Ours	0.087	0.316	0.037	0.935	0.992	0.998

Method	AbsRel↓	RMSE↓	log10↓	δ<1.25↑	δ<1.25²↑	δ<1.25³↑
Eigen^[20]	0.158	0.641	-	0.769	0.950	0.988
DORN^[25]	0.115	0.509	0.051	0.828	0.965	0.992
BTS^[26]	0.110	0.392	0.047	0.885	0.978	0.994
TransDepth^[27]	0.106	0.365	0.045	0.900	0.983	0.996
Adabins^[11]	0.103	0.364	0.044	0.903	0.984	0.997
DAV^[35]	0.108	0.412	-	0.882	0.980	0.996
VNL^[36]	0.108	0.416	0.048	0.875	0.976	0.994
TEDepth^[29]	0.100	0.349	0.043	0.907	0.987	0.998
P3Depth^[30]	0.104	0.356	0.043	0.898	0.981	0.996
NENet^[28]	0.100	0.349	0.043	0.907	0.987	0.998
NeWCRFs^[31]	0.095	0.334	0.041	0.922	0.992	0.998
DepthFormer^[9]	0.096	0.339	0.041	0.921	0.989	0.998
PixelFormer^[12]	0.090	0.322	0.039	0.929	0.991	0.998
GDM-Depth^[37]	0.113	0.439	0.049	0.872	0.972	0.993
Ours	0.087	0.316	0.037	0.935	0.992	0.998

Dataset	Method	Abs Rel ↓	RMSE ↓	Sq Rel ↓	δ<1.25↑
KITTI	B	0.060	2.430	0.180	0.960
	B+E	0.056	2.239	0.163	0.967
	B+C	0.054	2.225	0.161	0.971
	B+ E+C	0.056	2.205	0.158	0.970
	B+E+C+F	0.050	2.160	0.152	0.974
	B+ E+C+F+M (Ours)	0.048	2.094	0.147	0.980
NYU	B	0.103	0.342	0.054	0.910
	B+E	0.098	0.333	0.050	0.917
	B+C	0.095	0.334	0.047	0.922
	B+ E+C	0.094	0.331	0.047	0.923
	B+E+C+F	0.091	0.325	0.044	0.929
	B+ E+C+F+M (Ours)	0.087	0.316	0.040	0.934

融合层次化双流注意力的单目深度估计方法

Monocular depth estimation method with hierarchical dual-stream attention

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 37

相关文章 15

编辑推荐

Metrics

本文评价

方法	Params/M	FLOPs/G	GPU Memory (Allocated/GB)
NeWCRFs^[31]	270.33	395.56	1.02
Pixelformer^[12]	258.25	385.07	1.02
本文方法	229.70	349.22	0.87

[1]	李秀梅, 周正鑫, 孙军梅. 一种定位分支辅助的多任务协同图像伪造检测模型[J]. 图学学报, 2026, 47(3): 524-533.
[2]	李景涛, 封筠, 赵志宏. 适配器微调SAM与低频融合的遥感图像语义分割[J]. 图学学报, 2026, 47(3): 543-552.
[3]	卢德辉, 宋琢, 黄志超, 田时雨, 李慧敏, 田茂, 邓逸川. 基于TrueSkill排序与深度学习的绿色工地主观视觉感知预测[J]. 图学学报, 2026, 47(3): 641-652.
[4]	房友江, 王世豪, 张亮, 段可然, 刘越, 魏小鹏, 杨鑫. 基于图拓扑特征提取的跨模态一致性检测方法[J]. 图学学报, 2026, 47(2): 286-295.
[5]	闫康, 曾理, 顾晓清. 基于跨域结构化深度字典学习的图像分类方法[J]. 图学学报, 2026, 47(2): 341-350.
[6]	庞敏, 李振堂, 张元, 崔晓康, 熊风光. 基于检索与变形技术的三维模型重构[J]. 图学学报, 2026, 47(2): 368-379.
[7]	董文益, 杨伟东, 唐冰慧, 王琦, 肖宏宇. 基于深度学习的肝脏局灶性病变检测方法综述[J]. 图学学报, 2026, 47(1): 1-16.
[8]	张行顺, 陈海永. 基于动态视觉传感器的航发叶片缺陷检测[J]. 图学学报, 2026, 47(1): 120-130.
[9]	翟永杰, 王紫萱, 张祯琪, 周迅琪, 王乾铭. 融合双重注意力与加权动态卷积的车辆损伤分类模型[J]. 图学学报, 2026, 47(1): 17-28.
[10]	潘宇轩, 金锐, 刘雨, 张琳. 基于生成模型的无监督多视点立体视觉网络[J]. 图学学报, 2026, 47(1): 29-38.
[11]	酒明远, 吴国伟, 宋旭光, 李书攀, 徐明亮. 基于不确定性引导的智能强化主动学习图像分类方法[J]. 图学学报, 2026, 47(1): 47-56.
[12]	向梦丽, 黄志勇, 佘雅丽, 丁妥君. 一种大视角变换场景下的图像匹配方法[J]. 图学学报, 2026, 47(1): 90-98.
[13]	杨彪, 王学, 官铮, 龙萍. BSD-YOLO：基于动态稀疏注意力与自适应检测头的小目标车辆检测方法[J]. 图学学报, 2026, 47(1): 99-110.
[14]	琚晨, 丁嘉欣, 王泽兴, 李广钊, 管振祥, 张常有. 面向有限元法的图神经网络形函数近似方法[J]. 图学学报, 2025, 46(6): 1161-1171.
[15]	易斌, 张立斌, 刘丹楹, 唐军, 方俊俊, 李雯琦. 基于AMTA-Net的卷制过程激光打孔通风率预测模型[J]. 图学学报, 2025, 46(6): 1224-1232.