Laser & Optoelectronics Progress, Volume. 62, Issue 8, 0811005(2025)
Lightweight Unsupervised Monocular Depth Estimation Framework Using Attention Mechanisms
Figures & Tables(16)
Fig. 2. Model architectures of cross-covariance Transformer and coordinate attention
Table 1. Details of the composition of the depth estimation network
View tableTable 1. Details of the composition of the depth estimation network
Depth encoder Depth decoder Stage name Network structure Output size Stage name Network structure Output size Stage 1 7×7， 64， stride 2 96×320 Stage 1 UP-Block 3×3， 32， stride 1upsample
concatenate
3×3， 32， stride 1
192×640 Stage 2 Max Pool 3×3 Max Pool， stride 2 48×160 CCT-Block 3×3， 64， stride 1 3×3， 64， stride 1 CCT CA-Block 3×3， 64， stride 1 BR-Block 1×1， 1， stride 1 3×3， 64， stride 1 upsample CA Sigmoid Stage 3 CCT-Block 3×3， 128， stride 2 24×80 Stage 2 UP-Block 3×3， 64， stride 1 96×320 3×3， 128， stride 1 upsample 1×1， 128， stride 2 concatenate CCT 3×3， 64， stride 1 CA-Block 3×3， 128， stride 1 BR-Block 1×1， 1， stride 1 3×3， 128， stride 1 upsample CA Sigmoid Stage 4 CCT-Block 3×3， 256， stride 2 12×40 Stage 3 UP-Block 3×3， 128， stride 1 48×160 3×3， 256， stride 1 upsample 1×1， 256， stride 2 concatenate CCT 3×3， 128， stride 1 CA-Block 3×3， 256， stride 1 BR-Block 1×1， 1， stride 1 3×3， 256， strode 1 upsample CA Sigmoid
Table 2. Quantitative results on the KITTI dataset
View tableTable 2. Quantitative results on the KITTI dataset
Method Dataset Params /106↓ Abs_rel↓ Sq_rel↓ Rmse↓ Rmselog↓ a1↑ a2↑ a3↑ SfMLearner［ 7 ］M 16.5 0.183 1.595 6.709 0.270 0.734 0.902 0.959 Monodepth2-18［ 8 ］M 14.3 0.132 1.044 5.142 0.210 0.845 0.948 0.977 Monodepth2-50［ 8 ］M 32.5 0.131 1.023 5.064 0.206 0.849 0.951 0.979 DeepMatchVO［ 9 ］M 32.8 0.156 1.309 5.730 0.236 0.797 0.929 0.969 SGDepth［ 25 ］M+Se 16.3 0.128 0.973 5.085 0.206 0.853 0.951 0.978 R-MSFM3［ 26 ］M 3.5 0.128 0.965 5.019 0.207 0.853 0.951 0.977 Lite-mono-tiny［ 18 ］M 2.2 0.125 0.935 4.986 0.204 0.853 0.950 0.978 Proposed method M 4.9 0.122 0.937 4.942 0.199 0.858 0.953 0.981 Proposed method （4-layers） M 15.8 0.121 0.935 4.945 0.197 0.858 0.954 0.981 Monodepth2-18［ 8 ］M† 14.3 0.115 0.903 4.863 0.193 0.877 0.959 0.981 R-MSFM3［ 26 ］M† 3.5 0.114 0.815 4.841 0.190 0.866 0.857 0.982 Lite-mono-tiny［ 18 ］M† 2.2 0.110 0.837 4.710 0.187 0.880 0.960 0.982 Proposed method M† 4.9 0.107 0.839 4.674 0.183 0.883 0.963 0.982
Table 3. Model parametric quantities and computational analysis
View tableTable 3. Model parametric quantities and computational analysis
Method Encoder Decoder Full moder Params /106 FLOP /109 Params /106 FLOP /109 Params /106 FLOP /109 Monodepth2-18［ 8 ］11.2 4.5 3.1 3.5 14.3 8.0 DeepMatchVO［ 9 ］29.4 11.7 3.4 3.7 32.8 15.4 Lite-mono-tiny［ 18 ］2.0 2.4 0.2 0.44 2.2 2.84 R-MSFM3［ 26 ］1.7 2.4 3.8 14.1 5.5 16.5 Proposed method 4.1 5.0 0.8 2.8 4.9 7.8 Proposed method（4-layers） 12.6 6.0 3.2 3.6 15.8 9.6
Table 4. Quantitative results on the Make3D dataset
View tableTable 4. Quantitative results on the Make3D dataset
Method Abs_rel Sq_rel Rmse Rmeslog a1 a2 a3 Monodepth2-18［ 8 ］0.322 3.589 7.417 0.178 0.483 0.757 0.855 DeepMatchVO［ 9 ］0.436 5.101 8.748 0.197 0.443 0.684 0.817 Lite-mono-tiny［ 18 ］0.318 3.214 7.232 0.164 0.543 0.790 0.894 Proposed method 0.307 3.195 7.158 0.160 0.559 0.801 0.913
Table 5. Quantitative results on the Cityscapes dataset
View tableTable 5. Quantitative results on the Cityscapes dataset
Method Abs_rel Sq_rel Rmse Rmeslog a1 a2 a3 Monodepth2-18［ 8 ］0.386 3.752 8.556 0.278 0.451 0.717 0.843 DeepMatchVO［ 9 ］0.617 5.235 9.735 0.362 0.324 0.569 0.660 Lite-mono-tiny［ 18 ］0.406 3.992 8.752 0.292 0.377 0.654 0.735 Proposed method 0.360 3.711 7.873 0.274 0.482 0.734 0.868
Table 6. Quantitative results on the NUYDepth-v2 dataset
View tableTable 6. Quantitative results on the NUYDepth-v2 dataset
Method Abs_rel Sq_rel Rmse Rmeslog a1 a2 a3 Monodepth2-18［ 8 ］0.517 1.971 2.228 0.530 0.349 0.610 0.788 DeepMatchVO［ 9 ］0.682 4.676 4.317 0.824 0.186 0.355 0.537 Lite-mono-tiny［ 18 ］0.745 3.383 2.547 0.754 0.214 0.418 0.603 Proposed method 0.419 0.984 1.617 0.468 0.381 0.665 0.840
Table 7. Results of ablation experiments
View tableTable 7. Results of ablation experiments
Method Abs_rel Sq_rel Rmse Rmselog a1↑ a2↑ a3↑ Base/Monodepth2 0.132 1.044 5.142 0.210 0.845 0.948 0.977 Base + CCT-Block 0.128 0.978 5.013 0.205 0.856 0.952 0.979 Base + CA-Block 0.128 1.102 5.149 0.207 0.855 0.847 0.977 Base + CCT-Block + CA-Block 0.126 0.938 5.051 0.203 0.855 0.952 0.980 Base + CCT-Block + CA-Block + SURF 0.122 0.937 4.942 0.199 0.858 0.953 0.981 Method-1 0.136 1.115 5.233 0.213 0.839 0.947 0.976 Method-2 0.136 1.071 5.239 0.211 0.838 0.945 0.978
Table 8. Odometer trajectory errors and standard deviations on the KITTI-odometry dataset
View tableTable 8. Odometer trajectory errors and standard deviations on the KITTI-odometry dataset
Method sequence 01 sequence 07 sequence 09 sequence 10 Base/Monodepth2 0.046±0.020 0.024±0.014 0.061±0.032 0.039±0.025 Base + CCT-Block 0.030±0.016 0.014±0.009 0.016±0.009 0.016±0.011 Base + CA-Block 0.031±0.017 0.014±0.010 0.018±0.011 0.017±0.012 Base + CCT-Block + CA-Block 0.025±0.017 0.013±0.008 0.016±0.008 0.014±0.010 Base + CCT-Block + CA-Block + SURF 0.017±0.010 0.010±0.007 0.011±0.004 0.012±0.008
Tools
Get Citation
Copy Citation Text
Xiyu Li, Yilihamu Yaermaimaiti, Lirong Xie, Shuoqi Cheng. Lightweight Unsupervised Monocular Depth Estimation Framework Using Attention Mechanisms[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0811005
Paper Information
Category: Imaging Systems
Received: Jul. 15, 2024
Accepted: Oct. 12, 2024
Published Online: Apr. 2, 2025
The Author Email: Yilihamu Yaermaimaiti (65891080@qq.com)
CSTR:32186.14.LOP241688
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20