xusyy/swinmono

SwinMono: Swin Transformer-based Monocular Depth Estimation on SeasonDepth

Dataset

We perform our experiments on [SeasonDepth].

Environment

Tested on

python==3.7.7
torch==1.11.0
h5py==3.6.0
scipy==1.7.3
opencv-python==4.5.5
mmcv==1.5.0
timm=0.5.4
albumentations=1.1.0
tensorboardX==2.5
gdown==4.4.0

You can install above package with

$ pip install -r requirements.txt

Test

Download our pretrained model [link] and put it in './ckpt/' folder.

$ python ./code/test.py --data_path <dir_to_data> --save_eval_pngs

Train

Download the pretrained model 'swin_base_patch4_window7_224_22k.pth' provided by Swin-Transformer and put it in './ckpt/' folder.

$ python ./code/train.py --data_path <dir_to_data> --exp_name train --save_model --save_result

License

For non-commercial purpose only (research, evaluation etc).

References

[1] Global-Local Path Networks for Monocular Depth Estimationwith Vertical CutDepth. [code]

[2] Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. [code]