Framework for the Complete Gaze Tracking Pipeline

The figure below shows a general representation of the camera-to-screen gaze tracking pipeline [1]. The webcam image is preprocessed to create a normalized image of the eyes and face, from left to right. These images are fed into a model, which predicts the 3D gaze vector. The predicted gaze vector can be projected onto the screen once the user’s head pose is known.
This framework allows for the implementation of a real-time approach to predict the viewing position on the screen based only on the input image.

camera-to-screen gaze tracking pipeline

  1. pip install -r requirements.txt
  2. If necessary, calibrate the camera using the provided interactive script python, see Camera Calibration by OpenCV.
  3. For higher accuracy, it is also advisable to calibrate the position of the screen as described by Takahashiet al., which provide an OpenCV and matlab implementation.
  4. To make reliable predictions, the proposed model needs to be specially calibration for each user. A software is provided to collect this calibration data.
  5. Train a model or download a pretrained model.
  6. If all previous steps are fulfilled, python --calibration_matrix_path=./calibration_matrix.yaml --model_path=./p00.ckpt can be executed and a “red laser pointer” should be visible on the screen. also provides multiple visualization options like:
    1. --visualize_preprocessing to visualize the preprocessed images
    2. --visualize_laser_pointer to show the gaze point the person is looking at on the screen like a red laserpointer dot, see the right monitor on the image below
    3. --visualize_3d to visualize the head, the screen, and the gaze vector in a 3D scene, see left monitor on the image below


[1] Amogh Gudi, Xin Li, and Jan van Gemert, “Efficiency in real-time webcam gaze tracking”, in Computer Vision – ECCV 2020 Workshops – Glasgow, UK, August 23-28, 2020, Proceedings, Part I, Adrien Bartoli and Andrea Fusiello, Eds., ser. Lecture Notes in Computer Science, vol. 12535, Springer, 2020, pp. 529–543. DOI : 10.1007/978-3-030-66415-2_34. [Online]. Available:


View Github