This program discusses a pipeline to find the disparity map of two stereo images that contain specific objects, which are cars in this case. In this work, we build a pipeline that is based on previous well studied and known techniques and algorithms. Our pipeline consists of an Object-detection layer followed by an images matching box that feeds the resulting good matches to an 8-point algorithm that approximates the Fundamental matrix of the camera used in this work; at the end we build our disparity map over multiple correlation metrics between images like Zero Mean Normalized Cross Correlation (ZNCC), Sum of Absolute Differences (SAD) and Sum of Square Differences (SSD). This work could be used for autonomous car projects that tries to estimate depth of objects in images. For more details, refer to our report in the repository.
For the requirements refer to the file requirements.txt.
Clone the repo using:
git clone https://github.com/AyhamOlk/depth_estimator.git
which will create a folder called depth_estimator. Then run the following commands in sequence:
cd depth_estimator
mkdir build
cd build
cmake ..
make
After these commands the following exectuable files will be created: approximate_F and disparity. To be able to achieve the depth map for certain images, make sure to place the correct paths in the detect.py and disparity.cpp. Finally the whole story is about typing the following commands:
python3 detect.py
./approximate_F
./disparity
The resulting depth map image will be located in the depth_estimator/runs/disparity folder under the name disparity.jpg.
We have also other file in the repository called pixel_SAD.cpp and correspondingly its executable file will be pixel_SAD. If you want to run the version with window of size 1 (i.e. movement pixel by pixel) where we applied SAD on 2 car images. To be able to see the results of this implementation it is enough to run the following commands:
python3 detect.py
./pixel_SAD
which will output you the disparity map (best result that we got), but also will save it under the path depth_estimator/runs/disparity/disparity2.jpg.
In case of facing issues, do not hesitate to contact us: Ayham Olleik ayham.olleik@polytechnique.edu Khalig Aghakarimov khalig.aghakarimov@polytechnique.edu