YACHT: Yet Another Camera Hand calibration Tool

• Input data:

  • Arm poses
  • Camera images

• Output data:

  • Camera to arm end effector pose
  • Camera intrinsics
  • Target to arm base pose

Running the script

• ROS not required

• Install dependencies:

  pip install -r requirements.txt
  

• Edit config.yaml:

  • Set data-folder path
  • Set chessboard target size

• Input data format:

  • your_data_folder/images — camera to arm calibration images named like 0001.png (order matters)
  • your_data_folder/arm_poses.npy — [N, 4, 4] numpy array of poses

• Run:

  python yacht.py
  

• Loading output data:

with h5py.File("your_data_folder/result.h5", "r") as f:
    cam_to_arm_pose = f["cam_to_arm_pose"][:] # [4, 4] pose matrix
    camera_matrix = f["camera_matrix"][:] # [3, 3] camera matrix
    distortion_coeffs = f["distortion_coefficients"][:] # vector sized 5
    target_to_base_pose = f["target_to_base_pose"][:] # [4, 4] pose matrix

Output & Checks

• Calibration RMS error should be low — otherwise rerecord the data

• Validate cam-to-arm pose and error — rerecord the data/check the frames if it's high

• With visualize-2D: → your_data_folder/visualization/ contains calibration target frame reprojected into images

• With visualize-3D:

  • First view — camera poses w.r.t. target
  
  • Second view — camera, arm, target poses in global frame. Target frames (frames in front of cameras) should cluster tightly. If spread out — error is high, add more diverse data
  

Example dataset

Located in example_dataset folder

Author's note

Made by Nikolai Goncharov.