MiDaS

calendar Feb 6, 2023 · 1 min read · ai video depth map  ·
Share on: linkedin copy

GitHub Repository

During installation, I ran into an issue where the CUDA package wasn't found. Had to modify environment.yaml to:

 1name: midas-py310
 2channels:
 3  - pytorch
 4  - defaults
 5dependencies:
 6  - nvidia::cuda-toolkit=11.7.0
 7  - python=3.10.8
 8  - pytorch::pytorch=1.13.0
 9  - torchvision=0.14.0
10  - pip=22.3.1
11  - numpy=1.23.4
12  - pip:
13    - opencv-python==4.6.0.66
14    - imutils==0.5.4
15    - timm==0.6.12
16    - einops==0.6.0

Commands that were helpful for troubleshooting CUDA:

1conda list env
2conda env remove -n midas-py310
3python -m torch.utils.collect_env
4nvidia-smi
5conda install cudatoolkit

1conda install -c "nvidia/label/cuda-11.7.0" cuda-toolkit

Activate the Conda environment

1conda activate midas-py310

Run MiDaS

From the Conda Shell, cd to the MiDaS directory
1cd C:\Users\trima\MiDaS

Place the image frames you would like to process in the "input" directory and run one of the following commands:

1# dpt_beit_large_512
2python run.py --model_type dpt_beit_large_512 --input_path input --output_path output --grayscale --optimize
3# dpt_swin2_large_384
4python run.py --model_type dpt_swin2_large_384 --input_path input --output_path output --grayscale --optimize
5# dpt_swin2_tiny_256
6python run.py --model_type dpt_swin2_tiny_256 --input_path input --output_path output --grayscale --optimize

For "inferno" color mapping, omit the --grayscale flag.