doc-vu / monodepth2

Monocular depth estimation from a single image

Monodepth2

Based on the original version of Monodepth2, We added the test_cam.py to test Monodepth2 on image stream using a webcam.

⚙️ Setup

You can refer to the README from the original project to setup the required environment or just quickly go through the following steps:

Assuming a fresh Anaconda distribution, you can install the dependencies with:

conda install pytorch=0.4.1 torchvision=0.2.1 -c pytorch
pip install tensorboardX==1.4
conda install opencv=3.3.1   # just needed for evaluation and display the result

🔧 Usage

You can predict depth for a single image with:

python test_simple.py --image_path assets/test_image.jpg --model_name mono+stereo_640x192

Or you can predict depth using image stream with a webcam:

python test_cam.py

# The arguments can be added to run the above command is listed below
--model_name
# default="mono+stereo_640x192"
# choices=["mono_640x192",
# "stereo_640x192",
# "mono+stereo_640x192",
# "mono_no_pt_640x192",
# "stereo_no_pt_640x192",
# "mono+stereo_no_pt_640x192",
# "mono_1024x320",
# "stereo_1024x320",
# "mono+stereo_1024x320"]
--no_cuda
# if set, disables CUDA
--webcam
# default=0
# change to the desired by changing the value
--no_process
# if set, displays image in current process, might improve performance on machines without a GPU
--no_blend
# if set, does not display blended image
--no_display
# if set, does not display images, only prints fps