conda env create -f depthestimate_env.yaml
conda activate depthestimate_env
for mac m1 use depthestimate_env_mac_cpu.yaml
Training your model
python train.py --model MONODEPTH2 --conf configs/model_config.cfg
To run in background
nohup python -u train.py --model MONODEPTH2 > output.log &
Model | Additions | Link |
---|---|---|
CAMLESS | Learnable Camera Intrinsics | link |
ESPCN | Using ESPCN for Upsampling | link |
CAMLESS_WEATHER_AUGMENTATION | CAMLESS with weather augmentation | link |
MASKCAMLESS | Semantic segmentation suggestion from pretrained MASK-RCNN Model + CAMLESS | link |
MASKCAMLESS_V2 | MASKCAMLESS + skipping loss adjustment for Smoothness loss | link |
MASKCAMLESS_ESPCN | Mask R-CNN + CAMLESS + ESPCN | link |
MASKCAMLESS_ESPCN_WEATHER | MASKCAMLESS_ESPCN + weather augmentation | link |
MASKCAMLESS_ESPCN_V2 | MASKCAMLESS_ESPCN+ skipping loss adjustment for Smoothness loss | link |
Implementation | a1 | a2 | a3 | abs_rel | log_rms | rms | sq_rel |
---|---|---|---|---|---|---|---|
MonoDepth2 [6] | 0.877 | 0.959 | 0.981 | 0.115 | 0.193 | 4.863 | 0.903 |
CamLess[10] | 0.891 | 0.964 | 0.983 | 0.106 | 0.182 | 4.482 | 0.75 |
Ours - Monodepth2 + Mask R-CNN | 0.9008 | 0.9684 | 0.9872 | 0.1117 | 0.1886 | 3.977 | 0.5114 |
Ours - MonoDepth2 + Mask R-CNN + ESPCN | 0.8403 | 0.9651 | 0.9858 | 0.1214 | 0.205 | 4.096 | 0.6251 |
Ours - MonoDepth2 + CamLess | 0.8629 | 0.9542 | 0.98 | 0.1186 | 0.2103 | 4.737 | 0.7843 |
Ours - MonoDepth2 + CamLess+Weather Augmentation | 0.8704 | 0.9582 | 0.9789 | 0.1223 | 0.2016 | 4.934 | 0.9271 |
Ours - MonoDepth2 + Mask R-CNN + CamLess | 0.9148 | 0.9685 | 0.9832 | 0.0996 | 0.1887 | 4.25 | 0.5722 |
Ours - MonoDepth2 + Mask R-CNN + CamLess (Adjusted Loss) | 0.879 | 0.9699 | 0.9876 | 0.111 | 0.177 | 3.959 | 0.5079 |
Ours - MonoDepth2 + Mask R-CNN + ESPCN + CamLess | 0.9105 | 0.9637 | 0.9814 | 0.0956 | 0.1858 | 3.746 | 0.4868 |
Ours - MonoDepth2 + Mask R-CNN + ESPCN + CamLess (Adjusted Loss) | 0.8854 | 0.9621 | 0.9842 | 0.1166 | 0.1884 | 3.485 | 0.4793 |