With the rising demand for edge devices, lightweight self-supervised monocular depth estimation methods have continued to evolve; however, existing methods generally suffer from poor accuracy. We find that the root cause may lies in the failure of conventional training regimes to fully exploit the inherent potential of lightweight models. To address this, we propose a novel three-phase training strategy for lightweight self-supervised monocular depth estimation, termed Signal Scaffolding (SS). Without introducing any additional dataset or parameters, we apply SS to ReForgeDepth, a lightweight network of our own design. Experiments on the KITTI and Cityscapes datasets demonstrate that our method achieves state-of-the-art performance in the accuracy-efficiency trade-off and running in real time on edge devices. These results suggest that lightweight models can be improved further within their existing capacity, and that the central question may be less about how parameters are allocated than about how to unlock this latent potential through training.
Figures & tables
Figure 1: Performance and qualitative comparison. (a) Qualitative depth predictions: the proposed ReForgeDepth recovers finer boundary details than representative lightweight methods while using the fewest parameters. (b) Accuracy-cost trade-off on KITTI, where ReForgeDepth sits closer to the upper right corner.
Figure 2: (a) Overall Architecture of the ReForgeDepth. The proposed architecture comprises three innovative modules: GMD, and DCGI, with detailed implementations illustrated in Figure 3 . (b) The proposed Signal Scaffolding training strategy.
Figure 3: Structures of the proposed GMD and DCGI module.
Figure 4: PoseNet is employed to estimate the poses between adjacent frames.
Figure 5: Qualitative results on the KITTI Eigen test split. Zoom-in views of the regions marked by colored boxes. Monodepth2 [ 17 ] and Lite-Mono [ 59 ] blur object boundaries and lose fine details, whereas R-MSFM-GC [ 64 ] and PuriLight [ 4 ] oversmooth fine structures. ReForgeDepth, in contrast, preserves sharp boundaries and fine details.
Method
Year
Train
Depth Error ↓
Depth Accuracy ↑
Model Size ↓ Params.
Abs Rel
Sq Rel
RMSE
RMSE log
δ<1.25
δ<1.252
δ<1.253
KITTI(640 × 192)
Lightweight ( < 5M)
Lite-Mono-tiny [ 59 ]
2023
M
0.110
0.837
4.710
0.187
0.880
0.960
0.982
2.2M
Lite-Mono-small [ 59 ]
2023
M
0.110
0.802
4.671
0.186
0.879
0.961
0.982
2.5M
Lite-Mono [ 59 ]
2023
M
0.107
0.765
4.561
0.183
0.886
0.963
0.983
3.1M
Table 1: Quantitative results. We compare ReForgeDepth with representative methods on the KITTI Eigen split [ 7 ] . The best and second-best results in lightweight are highlighted in bold and underlined , respectively. M: KITTI monocular videos.
Method
Abs Rel ↓
Sq Rel ↓
RMSE ↓
RMSE log ↓
Params ↓
DDVO [ 47 ]
0.387
4.720
8.090
0.204
28.1M
Monodepth2 [ 17 ]
0.322
3.589
7.417
0.163
14.3M
R-MSFM6 [ 65 ]
0.334
3.285
7.212
0.169
3.8M
Lite-Mono-tiny [ 59 ]
0.318
3.217
7.234
0.164
2.2M
Flex-Small [ 66 ]
0.315
3.127
6.865
0.159
6.1M
ReForgeDepth(Ours)
0.311
3.094
6.995
0.160
1.5M
Table 2: Quantitative results. Comparison of ReForgeDepth with representative methods on the Make3D [ 43 ] dataset.
Method
Full Model ↓
Macbook Air (M2)
iPhone 12 (A14)
Params(M)
FLOPs(G)
Abs Rel
Speed (ms) ↓
FPS ↑
Speed (ms) ↓
FPS ↑
Monodepth2 [ 17 ]
14.3
8.0
0.115
21.36
47
30.47
33
R-MSFM3 [ 65 ]
3.5
16.5
0.114
21.61
46
47.53
21
Lite-Mono [ 59 ]
3.1
5.0
0.107
18.27
55
41.65
24
PuriLight [ 4 ]
2.7
7.1
0.106
37.26
27
60.64
16
ReForgeDepth
1.5
3.6
0.102
17.28
58
29.80
34
Table 3: Model complexity and inference speed. All models process 640 × 192 inputs. Params: number of parameters; FLOPs:floating-point operations; FPS: frames per second.
Abs Rel ↓
Sq Rel ↓
RMSE ↓
Method
w/o
w/
w/o
w/
w/o
w/
(a) Teacher: Lite-Mono-8M
Lite-Mono-tiny
0.110
0.106
0.837
0.799
4.710
4.574
Flex-Nano
0.110
0.108
0.794
0.789
4.678
4.627
Lite-Mono
0.107
0.105
0.765
0.747
4.561
4.511
ReForgeDepth (ours)
0.110
0.104
0.894
0.746
4.748
4.521
Table 4: Generalization of the Signal Scaffolding (SS) on the KITTI Eigen split. “w/o SS” denotes training with ImageNet initialization only, and “w/ SS” further applies Signal Scaffolding.
Weight η
Warm-up
Start ep.
Teacher
Abs Rel ↓
(a) Distillation weight, activated from epoch 0
0.1
✗
0
LiteMono-8M
0.106
0.1
✗
0
Flex-Large
0.104
0.05
✗
0
LiteMono-8M
0.107
0.05
✗
0
Flex-Large
0.105
0.025
✗
0
LiteMono-8M
0.108
Table 5: Sensitivity of the weight η and its schedule on the KITTI Eigen split. η is the coefficient of the distillation loss in Eq. ( 20 ); “Start ep.” denotes the epoch at which distillation is activated, and “Warm-up” indicates whether η increases gradually from zero.
Configuration
KITTI Eigen
Phase I
Phase II
Phase III
Abs Rel ↓
(a) Training-component ablation
✗
✗
Lite-Mono-8M
0.120
✗
✗
Flex-Large
0.147
✗
✓
✗
0.117
✗
✓
Lite-Mono-8M
0.115
Table 6: Ablation of training components and teacher choice. Phase I denotes ImageNet initialization, Phase II denotes photometric self-supervised training, and Phase III denotes teacher distillation. A teacher name indicates that Phase III is used; ✗indicates that it is omitted. Best Abs Rel is in bold .
Method
Year
Train
Depth Error ↓
Depth Accuracy ↑
Model Size ↓ Params.
Abs Rel
Sq Rel
RMSE
RMSE log
δ<1.25
δ<1.252
δ<1.253
416 × 128 (M)
Lightweight ( < 5M)
Flex-Nano [ 66 ]
2026
M
0.107
1.261
6.133
0.164
0.893
0.971
0.989
1.5M
PuriLight [ 4 ]
2025
M
0.099
1.014
5.892
0.153
0.901
0.976
0.992
2.7M
Ours
2026
M
0.091
0.918
5.590
0.146
0.912
0.978
0.992
1.5M
Table 7: Quantitative results. We compare ReForgeDepth with representative methods on the Cityscapes . The best and second-best results in lightweight are highlighted in bold and underlined , respectively. M: monocular videos, MS: monocular videos + semantic segmentation.
Abs Rel ↓
Sq Rel ↓
RMSE ↓
RMSE log ↓
Method
w/o
w/
w/o
w/
w/o
w/
w/o
w/
(a) Teacher: Lite-Mono-8M [ 59 ]
Lite-Mono-tiny [ 59 ]
0.110
0.106
0.837
0.799
4.710
4.574
0.187
0.182
Flex-Nano [ 66 ]
0.110
0.108
0.794
0.789
4.678
4.627
0.184
0.184
Lite-Mono [ 59 ]
0.107
0.105
0.765
0.747
4.561
4.511
0.183
0.181
ReForgeDepth (ours)
0.110
0.104
0.894
0.746
4.748
4.521
0.186
0.180
Table 8: Generalization of the Signal Scaffolding (SS) on the KITTI Eigen split. “w/o SS” denotes training with ImageNet initialization only, and “w/ SS” further applies Signal Scaffolding.
Method
Year
Train
Depth Error ↓
Depth Accuracy ↑
Model Size ↓ Params.
Abs Rel
Sq Rel
RMSE
RMSE log
δ<1.25
δ<1.252
δ<1.253
1024 × 320 (M ∗ )
Lightweight ( < 5M)
R-MSFM3 [ 65 ]
2021
M ∗
0.112
0.773
4.581
0.189
0.879
0.960
0.982
3.5M
R-MSFM6 [ 65 ]
2021
M ∗
0.108
0.748
4.470
0.185
0.889
0.963
0.982
3.8M
MViTDepth-tiny [ 12 ]
2023
M ∗
0.109
0.818
4.603
0.184
0.887
0.963
0.983
1.8M
Table 9: Quantitative results. We compare ReForgeDepth with representative methods on the KITTI Eigen split [ 7 ] . The best and second-best results in lightweight are highlighted in bold and underlined , respectively. M † : not pre-trained on ImageNet [ 6 ] , M: KITTI monocular videos, MS: monocular videos + semantic segmentation, M ∗ : input resolution of 1024×320 .
Figure 6: Qualitative results on the Make3D dataset. Validation on challenging scenes from the Make3D dataset. Compared with prior lightweight models, ReForgeDepth captures a wider receptive field and finer details, demonstrating superior generalization to unseen data.
Weight η
Warm-up
Start ep.
Teacher
Abs Rel ↓
Sq Rel ↓
RMSE ↓
RMSE log ↓
δ<1.25↑
δ<1.252↑
δ<1.253↑
(a) Distillation weight, activated from epoch 0
0.1
✗
0
LiteMono-8M [ 59 ]
0.106
0.795
4.603
0.183
0.887
0.962
0.983
0.1
✗
0
Flex-Large [ 66 ]
0.104
0.772
4.552
0.181
0.888
0.963
0.983
0.05
✗
0
LiteMono-8M [ 59 ]
0.107
0.820
4.641
0.184
0.886
0.962
0.982
0.05
✗
0
Flex-Large [ 66 ]
0.105
0.795
4.605
0.182
0.887
0.962
0.983
0.025
✗
0
LiteMono-8M [ 59 ]
0.108
0.841
4.675
0.184
0.885
0.962
0.982
Table 10: Sensitivity of the weight η and its schedule on the KITTI Eigen split. η is the coefficient of the distillation loss in Eq. ( 20 ); “Start ep.” denotes the epoch at which distillation is activated, and “Warm-up” indicates whether η increases gradually from zero.
Encoder configuration
Accuracy
Complexity
CDC
GMD
LGFI
DCGI
BR
Abs Rel ↓
Sq Rel ↓
RMSE ↓
RMSE log ↓
δ<1.25↑
δ<1.252↑
δ<1.253↑
Params (M) ↓
FLOPs (G) ↓
✓
0.124
0.995
4.995
0.199
0.860
0.954
0.979
2.67
4.403
✓
0.124
0.952
4.983
0.203
0.850
0.952
0.979
1.29
3.253
✓
✓
0.121
0.876
4.918
0.199
0.859
0.953
0.980
3.07
5.032
✓
✓
0.120
0.880
4.874
0.199
0.861
0.953
0.979
1.70
3.760
✓
✓
0.121
0.942
4.872
0.199
0.864
0.954
0.979
2.90
4.775
Table 11: Ablation study of the ReForgeDepth architecture on the KITTI Eigen split. CDC and LGFI are the encoder blocks and the feature-interaction module of Lite-Mono, respectively; GMD , DCGI , and BR are our contributions. All experiments were trained from scratch to highlight the impact of the architecture. Best accuracy is in bold .
Configuration
KITTI Eigen
Phase I
Phase II
Phase III
Abs Rel ↓
Sq Rel ↓
RMSE ↓
RMSE log ↓
δ<1.25↑
δ<1.252↑
δ<1.253↑
(a) Training strategies
✗
✗
Lite-Mono-8M [ 59 ]
0.120
0.890
4.887
0.198
0.860
0.954
0.980
✗
✗
Flex-Large [ 66 ]
0.147
1.105
5.444
0.219
0.807
0.939
0.977
✗
✓
✗
0.117
0.860
4.767
0.195
0.868
0.957
0.980
✗
✓
Lite-Mono-8M [ 59 ]
0.115
0.833
4.739
0.193
0.871
0.958
0.981
Table 12: Ablation of training components and teacher choice. Phase I denotes ImageNet initialization, Phase II denotes photometric self-supervised training, and Phase III denotes teacher distillation. A teacher name indicates that Phase III is used; ✗indicates that it is omitted. Best Abs Rel is in bold .
Self-supervised monocular depth estimation has emerged as an appealing solution to design lightweight and effective models for deployment on computationally constrained devices due to its reduced reliance on expensive depth sensors. By eliminating the need for ground-truth annotations and leveraging the simplicity of monocular camera setups, this approach facilitates cost-effective data collection and broad applicability across fields such as computer vision and robotics. A critical challenge is achieving resource-efficient neural networks without compromising the overall performance. State-of-the-art models generally adopt depth-wise convolutions and attention mechanisms; however, these functions often incur high energy costs and face compatibility issues in embedded environments. To address this, we propose XiDepth, a lightweight architecture based on the XiNet operator block, designed to enhance feature extraction while maintaining low computational complexity and energy demand. On the KITTI dataset, XiDepth achieves state-of-the-art performance with only 0.8M parameters. Tests on a Raspberry Pi 4 further confirm its suitability for real-world embedded applications, reducing FLOPs by 40% and energy consumption by 35% compared to leading methods.
Elena Izzo, Riccardo Toniolo, Lamberto Ballan
Department of Mathematics University of Padova Padova, Italy
Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place them far beyond the reach of embedded and mobile platforms. Lightweight alternatives exist, but have been developed almost exclusively within single-domain, self-supervised paradigms, failing silently under domain shift. We present ZipDepth, a compact monocular depth network that bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model over a large multi-domain training set. Comprising just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, taking a significant step towards the accuracy of foundation models with 50x more parameters.
Accurate monocular depth estimation serves as a core enabler for single camera scene understanding. However, existing self-supervised monocular depth estimation methods generally suffer from the bottleneck of inefficient cross-scale information interaction and difficulty in balancing local and global spatial modeling. In this paper, we propose CMambaDepth, a self-supervised framework that achieves efficient multi-scale feature fusion and fine-grained contextual modeling via channel-wise selective state propagation. Specifically, Bidirectional Channel Mamba (Bi-CMamba) aligns encoder features across scales and enables bidirectional information exchange among ordered scale groups. Unidirectional Channel Mamba (Uni-CMamba) progressively aggregates decoder features and retains fine-grained scale groups through a group selection mechanism for subsequent fusion. Furthermore, a Hybrid Attention Module (HAM) is introduced to combine large-kernel local context and Manhattan self-attention for complementary spatial modeling. Experimental results demonstrate that our method achieves highly competitive performance. Specifically, our model achieves an AbsRel of 0.094 and an RMSE of 4.156 on KITTI, and an AbsRel of 0.140 on DDAD. In the zero-shot cross-dataset generalization test on NYUv2, it attains an AbsRel of 0.232, outperforming the baseline RA-Depth by 7.2%.
Xuezhi Xiang, Jiayao Liu, Heqi Xiang +3
School of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China · Key Laboratory of Advanced Marine Communication and Information Technology, Harbin 150001, China · Department of Computer Science, University of Toronto, Toronto ON M5S 2E4, Canada +1