Table of Contents
Created by gh-md-toc
Authors:
- Elif Cansu YILDIZ
- Salih MARANGOZ
Please view this file with a markdown editor/reader (e.g. Typora) for a better experience.
- The project notebook can be viewed directly (without Github, only HTML) here and all experiments output here.
- The project report (PDF) can be downloaded/viewed here.
We added a desktop entry to the run viewer easily. Run these two commands to install the viewer for .pfm/.pgm
files. Open files with Open with Other Application
and select PFS Viewer
.
$ sudo apt install pfstools pfsview
$ tee $HOME/.local/share/applications/pfs_viewer.desktop << END
[Desktop Entry]
Exec=sh -c "/usr/bin/pfsin %f | /usr/bin/pfsview"
Name=PFS Viewer
Terminal=false
Type=Application
END
NOTE: We added venv_for_webp
just in case something changes in the future.
Unfortunately there was a problem with webp and PIL connection. We solved this problem temporarily by installing the package to a virtual environment and appending its modules path to the PATH environment variable. Run these terminal commands to create the virtual environment:
# If you want to install the virtual environment by yourself;
# Remove 'venv_for_webp' folder by hand please. We didn't put here rm -rf command because it can be dangerous...
$ pip3 install virtualenv
$ virtualenv venv_for_webp
$ source venv_for_webp/bin/activate
$ pip3 install webp
Run in jupyter notebok for testing:
import os, sys
import numpy as np
local_pip_path = os.path.join(os.getcwd(), "venv_for_webp/lib/python3.8/site-packages")
sys.path.append(local_pip_path)
import webp
test_image = webp.load_image("test.webp", mode="RGB")
print("Test Image Shape:", np.array(test_image).shape)
test_image
We added environment.yml
for Anaconda installation. We had to use PyTorch-Nightly to utilize RTX3090 in the laboratory. Specified PyTorch-Nightly version may not be available so try using version 1.10.0
or 1.9.1
.
We described where to put dataset folder in project_notebook.ipynb
.
Starting with default settings may show less data. Try with increasing samples:
$ cd run
$ tensorboard --samples_per_plugin="scalar=10000,images=200" --logdir .
We also experimented with OpenCV disparity estimation, but the results was of course not good...
Source: https://github.com/aliyasineser/stereoDepth/blob/master/stereo_depth.py
Input (Left Image) (Right image not showed) | Output (Disparity Map) |
---|---|
![]() |
![]() |