In this project, we present a deep learning based method to perform hand-eye calibration in the wild. Our method does not require any object of known geometry to be present in the scene. It estimates the hand-eye calibration parameters by directly making sense of the visual features present in the images taken from different view-points.
For more details, visit the project website.
- The quaternion convention followed in this repository is (w, x, y, z)
To setup the environment
git clone https://github.com/mohith-sakthivel/deep_hand-eye_calib.git
cd deep_hand-eye_calib
conda env create -f environment.yaml
conda activate hand-eye
To download and setup the dataset
wget http://roboimagedata2.compute.dtu.dk/data/MVS/Rectified.zip
unzip Rectified.zip -d data/DTU_MVS_2014
mv temp/camera_pose.json data/DTU_MVS_2014
To train the model
python -m deep_hand_eye.train