This repositry is majorly based on the depth estimation GitHub repository 'monodepth2' by nianticlabs. Link to original repo:- https://github.com/nianticlabs/monodepth2 . For perspective transformation i referred to the following GitHub repo. :-https://github.com/darylclimb/cvml_project/tree/master/projections/inverse_projection I tried to analyse the different models proposed by them and drawn comparision between them in the documentation (Check Documentation Folder). An accuracy based comparision is available on the original Repo.
I've added 2 files 'geometry_utils.py' and 'finalDemo.py' with some references to original repository. The 'finalDemo.py' can be used to test a single mono image to estimate the depth image ,setting the relative range of depth and plotting the 3-D point clouds using open3D library.
The finalDemo.py file can be used to test and get 3-D visualization of the single image. It takes 2 arguments that are image path and model name. Choices for model are the same as the original pre-trained models from monodepth2 : 1.mono_640x192 2.stereo_640x192 3.mono_1024x320 4.stereo_1024x320
Running the file:- Use the following command to run the finalDemo.py file:
python finalDemo.py --image_path assets/s3.png --model_name mono_640x192The output of the file will be a depth vs rgb image comparision and Open3D 3-D plot of the 2-D image.
- conda install pytorch=0.4.1 torchvision=0.2.1 -c pytorch
- pip install tensorboardX==1.4
- conda install opencv=3.3.1
- pip install open3d