Acta Optica Sinica (Online), Volume. 2, Issue 13, 1305001(2025)
A Survey of Monocular Depth Estimation Methods Based on Deep Learning
Figures & Tables(16)
Fig. 2. Workflow of monocular depth estimation based on deep learning
Fig. 3. An example framework of monocular depth estimation based on supervised learning[29]
Fig. 4. Schematic diagram of geometric relationships between image pairs
Table 1. Comparison of several active and passive depth acquisition methods
View tableView in ArticleTable 1. Comparison of several active and passive depth acquisition methods
Category Method Principle Advantage Disadvantage Active Time-of-flight （ToF） Compute distance based on light travel time Simple principle， unaffected by ambient light High measurement cost Structured light Estimate depth using phase information of deformed light Effective for complex surfaces High measurement cost， sensitive to ambient light Passive Stereo depth estimation Disparity matching from stereo images Unrestricted by training data Struggle in low-texture scenes， poor performance at long distances Multi-view depth estimation Extract disparity information from multiple images Provide richer viewpoint information， effectively handle occlusion High computational complexity， poor real-time performance Monocular depth estimation Estimate depth from a single image Require only one camera， low cost， easy to integrate Rely on learning models， sensitive to training data and scene variations
Table 2. Comparison between this work and existing review papers
View tableView in ArticleTable 2. Comparison between this work and existing review papers
Ref. Year Supervision classification Recent work Method comparison Quantitative analysis Application overview ［ 7 ］2018 Supervised only No No Yes No ［ 8 ］2019 None No No No No ［ 10 ］2019 Supervised， unsupervised No No Yes No ［ 11 ］2020 Supervised， self-supervised， semi-supervised No Yes Yes Yes ［ 9 ］2021 Supervised， unsupervised No Yes Yes Yes ［ 12 ］2022 None No No Yes No ［ 13 ］2022 Supervised， unsupervised No No Yes No ［ 18 ］2022 Supervised， unsupervised， semi-supervised No Yes Limited No ［ 14 ］2022 Supervised， unsupervised， semi-supervised No No Yes No ［ 15 ］2023 None Yes No Yes No ［ 16 ］2024 Unsupervised only No Yes Yes Yes ［ 17 ］2024 Unsupervised only Yes Yes Yes No ［ 19 ］2024 Supervised， self-supervised， semi-supervised Yes Yes Yes Yes This work 2025 Supervised， unsupervised， semi-supervised Yes Yes Yes Yes
Table 3. Evolution of monocular depth estimation methods
View tableView in ArticleTable 3. Evolution of monocular depth estimation methods
Category Period Method Advantage Limitation Depth cues-based Before 2009 Shadows， focus/defocus Low cost， simple inefficient， non-real-time Traditional machine learning-based 2009‒2014 Parametric/non-parametric Low computation Texture-dependent， non-real-time Deep learning-based After 2014 Supervised， unsupervised， semi-supervised High accuracy， multi-scene， real-time High computation， data-hungry
Table 4. Monocular depth estimation methods based on supervised learning
View tableView in ArticleTable 4. Monocular depth estimation methods based on supervised learning
Ref. Method Contribution Limitation ［ 29 ］Multi-scale depth network Preserve details without superpixels Scale estimation errors ［ 31 ］ResNet， BerHu loss Address deep network issues Limited generalization ［ 33 ］Skip connections Prevent inter-layer info loss ‒ ［ 35 ］Laplacian pyramid Adaptive multi-scale fusion Lack global consistency ［ 39 ］Transformer extractor Global receptive field Weak local connectivity ［ 40 ］Multi-Transformer blocks Integrate local‒global info， restores details Miss long-range depth relations ［ 41 ］Transformer + CNN Enhances local features High computation ［ 42 ］Cross-distillation Leverage long-range‒local depth， reduces cost ‒ ［ 43 ］Depth gradient refinement Improve depth gradient continuity ‒ ［ 46 ］CRF superpixel refinement Integrate color， texture， position Struggle with complex scenes ［ 47 ］Fully connected CRF Fuse multiple features High computation ［ 48 ］Semantic integration More accurate far-depth estimation Limited receptive field ［ 49 ］Transformer + CNN Enhance long-range depth modeling ‒ ［ 50 ］Scene-aware estimation Depth at semantic-segment level Lack structural consistency ［ 51 ］Edge-assisted depth High-detail maps Boundary artifacts ［ 52 ］PatchFusion Merge coarse‒global and fine‒local depth Need real-time optimization ［ 54 ］Adaptive depth distribution Lightweight， suited for indoors ‒ ［ 55 ］BerHu sum loss Multi-scale fusion Unstable with outliers ［ 56 ］Hybrid loss Noise robustness Focus only on loss optimization ［ 57 ］3D geometric loss Faster computation， easy integration ‒ ［ 59 ］Relative depth Introduce related dataset View synthesis artifacts ［ 60 ］Structural ranking loss Reduce artifacts Inaccurate absolute depth ［ 61 ］Two-stage relative‒absolute depth Improve absolute depth ‒ ［ 62 ］Depth classification Lower model complexity Higher uncertainty at greater depth ［ 63 ］Progressive binning Reduce uncertainty Lack mathematical reliability ［ 64 ］Discrete classification Ensure depth probability distribution Ignore depth ordering ［ 65 ］Constrained ordinal regression Maintain depth order Inefficient ［ 66 ］Adaptive depth binning Eliminate quantization artifacts Fixed bin count ［ 69 ］LocalBins Iterative grouping refinement High complexity ［ 70 ］IEBins Maintain stable bin counts Weak detail capture ［ 71 ］CaBins Improve detail accuracy ‒ ［ 73 ］GAN-based First to apply GAN in depth estimation Poor texture recovery ［ 74 ］UNet + multi-skip Recover low-dim textures Lack conditional guidance ［ 75 ］cGAN Enhance depth estimation and reconstruction Sensitive to noisy conditions
Table 5. Monocular depth estimation methods based on unsupervised learning
View tableView in ArticleTable 5. Monocular depth estimation methods based on unsupervised learning
Ref. Method Key contribution Limitation ［ 84 ］Gradient smoothness & photometric loss First unsupervised training with image pairs Low accuracy， occlusion issues， local optima ［ 85 ］Left-right consistency Solve occlusion problem Texture-copying artifacts ［ 86 ］Pyramid processing Mitigate multi-scale object impact Depth holes in low-scale disparity maps ［ 87 ］Local plane parameter module Address local differentiability Limited dataset usage ［ 88 ］Dense feature fusion Smoother depth maps with clearer edges Limited dataset usage ［ 92 ］Coupled depth-pose network First video-based unsupervised framework Struggle with dynamics and occlusion ［ 93 ］Multi-scale appearance matching loss Reduce depth artifacts Require known camera intrinsics ［ 95 ］Geometry-aware self-discovered mask Solve scale inconsistency & occlusion Lack absolute scale factor ［ 96 ］Overlapping & blank masks Improve accuracy Poor in dynamic scenes ［ 97 ］VO joint estimation Enhance dynamic object estimation Scale ambiguity ［ 98 ］Depth-pose consistency loss & attention Reduce scale ambiguity Complex model ［ 100 ］3D geometric constraints No need for camera parameters Ignore moving objects ［ 101 ］Minimal instance photometric residual Remove dynamic object influence Depth detail improvement needed ［ 102 ］Occlusion-aware feature fusion Integrate spatial context ‒ ［ 104 ］Direct VO as pose predictor Solve scale ambiguity Poor for deformable objects ［ 106 ］Competitive collaboration framework Integrate multiple tasks Scale inconsistency ［ 107 ］Global loss constraint Solve scale inconsistency ‒ ［ 108 ］Cross-view completion Improve generalization High computation cost ［ 109 ］Shared feature pyramid Reduce network parameters Ignore sequence context ［ 110 ］ConvLSTM Utilize sequence context ‒ ［ 111 ］3D packing-unpacking blocks Preserve fine details Depend on accurate speed measurements ［ 112 ］Recurrent spatiotemporal network Enhance self-motion estimation High computation cost ［ 114 ］Joint framework （GeoNet） Reduce occlusion， blur， and dynamic object effects Optical flow depends on depth & pose networks ［ 115 ］Independent optical flow network Prevent error propagation Occlusion issues ［ 116 ］Multi-task joint estimation Mitigate occlusion Limited scene flow accuracy ［ 117 ］Independent motion estimation network More accurate scene flow Complex structure ［ 118 ］Rigid vs. full scene flow Reduce dynamic object influence Limited by brightness constancy assumption ［ 121 ］Semantic-guided depth （SGDepth） Reduce dynamic object impact Require known intrinsics ［ 123 ］Pixel-level semantic priors Improve depth consistency Limited receptive field ［ 124 ］Shared encoder with multi-task feature extraction Enhance weak-texture estimation ‒ ［ 126 ］Neighboring frames as auxiliary input No need for calibrated videos Sensitive to noise & local minima ［ 127 ］Self-cross-attention layers Improve feature matching Ignore dynamic objects ［ 128 ］Dynamic object motion disentanglement Reduce dynamic object effects Struggle in low-texture scenes ［ 129 ］Learnable PatchMatch Improve performance in low-texture or brightness variations Underutilize multi-view geometry ［ 130 ］Mono-depth as geometric prior for multi-view cost volume Robust under multi-view ambiguity ‒ ［ 132 ］GAN-based depth estimation First GAN-based unsupervised approach Limited accuracy ［ 133 ］GAN for depth & pose Improve accuracy Fail on dynamic objects ［ 134 ］GAN with temporal correlation Solve depth blur & pose inaccuracy Affected by occlusion & view changes ［ 135 ］Boolean masking Mitigate occlusion & view changes Ineffective in low-light ［ 136 ］Adversarial domain adaptation Enhance low-light performance Struggle with highlights & shadows ［ 137 ］Domain separation network Handle illumination changes Poor for single-day‒night scenes ［ 138 ］Multi-scale GAN Improve depth resolution High computation cost ［ 139 ］Diffusion-based model Enhance network stability ‒ ［ 140 ］Lightweight model Significantly reduce computation ‒
Table 6. Semi-supervised monocular depth estimation methods
View tableView in ArticleTable 6. Semi-supervised monocular depth estimation methods
Ref. Method Main contribution Limitation ［ 147 ］Sparse depth supervision Combine supervised and unsupervised constraints Limited accuracy ［ 148 ］Left-right consistency loss Improve accuracy Rely on a large amount of labeled disparity data ［ 149 ］Geometric constraints Reduce dependence on labeled data Efficiency needs improvement ［ 150 ］Student-teacher strategy Enhance training ‒ ［ 151 ］Motion cues + view changes Improve scale awareness Poor performance on dynamic objects ［ 152 ］Joint depth-semantic optimization First to integrate semantic information Limited dataset availability ［ 153 ］Symmetric depth estimation No need for paired depth data with semantic labels Domain shift issue ［ 154 ］Dual-branch framework Avoid data collection challenges Domain shift issue ［ 155 ］Domain-aware adversarial learning Mitigate cross-dataset differences Does not fully utilize global context ［ 156 ］Symbiotic Transformer + NearFarMix Integrate global and local information， reduce overfitting ‒ ［ 157 ］Cross-channel attention module Enhance cross-task channel interactions ‒ ［ 158 ］Manhattan world assumption Ensure real-time performance Limited applicability in complex environments ［ 159 ］Dual-discriminator GAN First GAN-based semi-supervised depth estimation model Limited accuracy ［ 161 ］Stacked recurrent GANs Reduce errors ‒
Table 7. Evaluation metrics
View tableView in ArticleTable 7. Evaluation metrics
Metric Formula Description RMSE Measure absolute error， and it is sensitive to outliers REL Log transformation reduces the influence of outliers Abs Measure relative error， but differences may not be significant Sq Squaring increases the error difference Acc Reflect prediction accuracy， where
Table 8. Quantitative evaluation results on the KITTI dataset
View tableView in ArticleTable 8. Quantitative evaluation results on the KITTI dataset
Ref. Supervision type RMSE REL Abs Sq Acc ［ 29 ］Supervised 7.156 0.270 0.190 1.515 0.692 0.899 0.967 ［ 35 ］Supervised 3.802 0.151 0.090 0.546 0.902 0.972 0.990 ［ 39 ］Supervised 2.573 0.092 0.062 ‒ 0.959 0.995 0.999 ［ 42 ］Supervised 2.032 0.076 0.050 0.142 0.977 0.997 0.999 ［ 43 ］Supervised 2.041 ‒ 0.049 ‒ 0.979 ‒ ‒ ［ 61 ］Supervised ‒ ‒ 0.057 ‒ ‒ ‒ ‒ ［ 62 ］Supervised 3.605 0.187 0.107 ‒ 0.898 0.966 0.984 ［ 65 ］Supervised 0.094 ‒ 0.180 ‒ 0.894 ‒ ‒ ［ 66 ］Supervised 2.360 0.088 0.059 0.190 0.964 0.995 0.999 ［ 71 ］Supervised 2.322 0.088 0.057 0.186 0.964 0.995 0.999 ［ 74 ］Supervised 2.702 0.116 0.072 0.325 0.951 0.989 0.996 ［ 84 ］Unsupervised 5.104 0.273 0.169 1.080 0.740 0.904 0.962 ［ 85 ］Unsupervised 5.927 0.247 0.148 1.344 0.803 0.922 0.964 ［ 88 ］Unsupervised 5.063 ‒ 0.122 0.939 0.850 0.947 0.976 ［ 92 ］Unsupervised 6.856 0.283 0.208 1.768 0.678 0.885 0.957 ［ 93 ］Unsupervised 4.863 0.193 0.115 0.903 0.877 0.959 0.981 ［ 95 ］Unsupervised 5.439 0.217 0.137 1.089 0.830 0.942 0.975 ［ 97 ］Unsupervised 4.813 0.192 0.117 0.863 0.871 0.959 0.982 ［ 98 ］Unsupervised 4.831 0.196 0.119 0.886 0.866 0.955 0.980 ［ 100 ］Unsupervised 6.220 0.250 0.163 1.240 0.762 0.916 0.968 ［ 104 ］Unsupervised 5.583 0.228 0.151 1.257 0.810 0.936 0.974 ［ 107 ］Unsupervised 5.515 0.211 0.148 1.098 0.819 0.934 0.973 ［ 108 ］Unsupervised 4.210 ‒ 0.098 ‒ 0.900 0.969 0.986 ［ 109 ］Unsupervised 4.727 ‒ 0.123 0.797 ‒ ‒ ‒ ［ 111 ］Unsupervised 4.601 0.189 0.111 0.785 0.878 0.960 0.982 ［ 112 ］Unsupervised 4.524 0.188 0.108 0.767 0.891 0.974 0.991 ［ 115 ］Unsupervised 5.507 0.223 0.150 1.124 0.806 0.933 0.973 ［ 117 ］Unsupervised 4.853 0.175 0.106 0.888 0.879 0.965 0.987 ［ 123 ］Unsupervised 4.839 0.196 0.114 0.846 0.853 0.953 0.982 ［ 126 ］Unsupervised 4.459 0.176 0.098 0.770 0.900 0.965 0.983 ［ 128 ］Unsupervised 4.458 0.175 0.096 0.720 0.897 0.964 0.984 ［ 130 ］Unsupervised 4.216 0.169 0.089 0.663 0.904 ‒ ‒ ［ 133 ］Unsupervised 5.448 0.216 0.150 1.141 0.808 0.939 0.975 ［ 134 ］Unsupervised 5.564 0.229 0.150 1.127 0.823 0.936 0.974 ［ 138 ］Unsupervised 5.632 0.234 0.144 0.148 0.795 0.927 0.971 ［ 147 ］Semi-supervised 3.518 0.179 0.108 0.595 0.875 0.964 0.988 ［ 148 ］Semi-supervised 3.464 0.126 0.078 0.417 0.923 0.984 0.995 ［ 150 ］Semi-supervised 3.162 0.162 0.091 0.505 0.901 0.969 0.986 ［ 151 ］Semi-supervised 3.349 0.118 0.074 0.355 0.930 ‒ ‒ ［ 152 ］Semi-supervised 6.526 0.222 0.143 2.161 0.850 0.939 0.972 ［ 154 ］Semi-supervised 4.361 0.217 0.116 0.956 0.848 0.948 0.974 ［ 156 ］Semi-supervised 0.289 ‒ 0.080 ‒ 0.948 ‒ ‒ ［ 158 ］Semi-supervised 3.032 0.112 0.070 0.314 0.945 0.974 0.990 ［ 159 ］Semi-supervised 4.195 0.170 ‒ 0.093 ‒ ‒ ‒
Table 9. Quantitative evaluation results on the NYU Depth-V2 dataset
View tableView in ArticleTable 9. Quantitative evaluation results on the NYU Depth-V2 dataset
Ref. Supervision type RMSE REL Abs Sq Acc ［ 29 ］Supervised 0.907 0.285 0.215 0.212 0.611 0.887 0.971 ［ 31 ］Supervised 0.573 0.195 0.127 ‒ 0.811 0.953 0.988 ［ 35 ］Supervised 0.514 ‒ 0.111 ‒ 0.878 0.977 0.994 ［ 39 ］Supervised 0.357 ‒ 0.110 ‒ 0.904 0.988 0.998 ［ 42 ］Supervised 0.316 ‒ 0.088 ‒ 0.933 0.992 0.998 ［ 43 ］Supervised 0.310 ‒ 0.086 ‒ 0.937 ‒ ‒ ［ 46 ］Supervised 0.821 ‒ 0.232 ‒ 0.621 0.886 0.968 ［ 48 ］Supervised 0.329 0.125 0.098 ‒ 0.917 0.983 0.996 ［ 50 ］Supervised 0.497 0.174 0.128 ‒ 0.845 0.966 0.990 ［ 54 ］Supervised 0.416 ‒ 0.121 ‒ 0.864 0.974 0.995 ［ 57 ］Supervised 0.552 ‒ 0.164 ‒ 0.768 0.940 0.984 ［ 59 ］Supervised 1.130 0.390 0.360 0.460 ‒ ‒ ‒ ［ 61 ］Supervised 0.277 ‒ 0.077 ‒ 0.953 0.995 0.999 ［ 62 ］Supervised 0.540 ‒ 0.141 ‒ 0.819 0.965 0.992 ［ 65 ］Supervised 0.297 ‒ 0.340 ‒ 0.667 ‒ ‒ ［ 66 ］Supervised 0.364 ‒ 0.103 ‒ 0.903 0.984 0.997 ［ 71 ］Supervised 0.401 ‒ 0.120 ‒ 0.866 0.978 0.996 ［ 73 ］Supervised 0.527 ‒ 0.134 0.103 0.822 0.971 0.993 ［ 74 ］Supervised 0.403 0.045 0.109 ‒ 0.871 0.979 0.993 ［ 98 ］Unsupervised 0.592 ‒ 0.165 ‒ 0.753 0.934 0.981 ［ 109 ］Unsupervised 0.377 ‒ 0.095 ‒ 0.905 0.980 0.995
Table 10. Quantitative evaluation results on the Cityscapes dataset
View tableView in ArticleTable 10. Quantitative evaluation results on the Cityscapes dataset
Ref. Supervision type RMSE REL Abs Sq Acc ［ 50 ］Supervised 6.917 0.414 0.227 3.800 0.801 0.913 0.950 ［ 100 ］Unsupervised 7.580 0.334 0.267 2.686 0.577 0.840 0.937 ［ 126 ］Unsupervised 6.223 0.170 0.114 1.193 0.875 0.967 0.989 ［ 128 ］Unsupervised 5.867 0.157 0.103 1.000 0.895 0.974 0.991 ［ 157 ］Semi-supervised 7.284 ‒ 0.142 1.553 0.824 0.956 0.987 ［ 158 ］Semi-supervised 7.163 0.411 0.289 3.884 0.661 0.854 0.935
Table 11. Quantitative evaluation results on the Make3D dataset
View tableView in ArticleTable 11. Quantitative evaluation results on the Make3D dataset
Ref. Supervision type RMSE REL Abs Sq ［ 31 ］Supervised 4.460 ‒ 0.176 ‒ ［ 46 ］Supervised 10.270 ‒ 0.279 ‒ ［ 74 ］Supervised 3.788 0.058 0.154 ‒ ［ 92 ］Unsupervised 10.470 0.478 0.383 5.321 ［ 93 ］Unsupervised 7.417 ‒ 0.322 3.589 ［ 97 ］Unsupervised 7.391 0.167 0.334 3.520 ［ 98 ］Unsupervised 7.609 ‒ 0.511 4.229 ［ 104 ］Unsupervised 8.090 0.204 0.387 4.720 ［ 109 ］Unsupervised 8.388 ‒ 0.367 4.267 ［ 115 ］Unsupervised 6.890 0.416 0.331 2.698 ［ 123 ］Unsupervised 7.646 ‒ 0.338 3.427 ［ 150 ］Semi-supervised 5.136 ‒ 0.107 0.265 ［ 154 ］Semi-supervised 6.497 0.158 0.402 4.181 ［ 161 ］Semi-supervised 66.370 ‒ 0.203 ‒
Table 12. Overview of depth estimation datasets
View tableView in ArticleTable 12. Overview of depth estimation datasets
Dataset Year Image type Sensor type Annotation type Scene type Image counts Resolution /（pixel×pixel） KITTI［ 162 ］2012 Stereo LiDAR Sparse Driving scene 4.4×104 1024×320 NYU Depth-V2［ 163 ］2012 Monocular Microsoft Kinect Dense Indoor 1449 640×480 Cityscapes［ 164 ］2016 Monocular Stereo Camera Disparity Urban driving 5×103 2048×1024 Make3D［ 165 ］2008 Monocular Laser Scanner Dense Outdoor scene 534 2272×1704 DIODE［ 166 ］2019 Stereo Laser Scanner Dense Indoor/outdoor 2.55×104 768×1024 Middlebury 2014［ 167 ］2014 Stereo DSLR Camera Dense Indoor detail 33 2872×1984 Driving Stereo［ 168 ］2019 Stereo LiDAR Sparse Driving scene 1.82×105 1762×800
Tools
Get Citation
Copy Citation Text
Yiquan Wu, Haobo Xie. A Survey of Monocular Depth Estimation Methods Based on Deep Learning[J]. Acta Optica Sinica (Online), 2025, 2(13): 1305001
Paper Information
Category: Optical Information Acquisition, Display, and Processing
Received: Apr. 25, 2025
Accepted: May. 15, 2025
Published Online: Jul. 7, 2025
The Author Email: Yiquan Wu (nuaaimage@163.com)
CSTR:32394.14.AOSOL250454
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20