LSC eye-tracker (Learning Single Camera Eye-Tracker)

This repo is for eye track research

We are in the process of preparing dataset for free public release now.

Data generation

• generate-dataset.py generate .npz data package for training from the MPIIFaceGaze dataset.
  • MPIIFaceGaze dataset arranged folder with different subjects and then different days' images data
  • We can read the directory from the p%02d.txt file to obtain the complete path for each images file and the corresponding gaze point on the screen
  • The screen size parameters inside the codes should be modified according to the values inside the screenSize.mat file in the Calibration folder
  • The saved .npz file contains two fields:
    1. faceData: nSample * rows * cols * channels;
    2. eyeTrackData: nSamples * v_grids * h_grids * 3 (relative x, y, and probability p)

$ python generate-dataset.py <folder (e.g., ./p00)> <save.npz>

Model training

• i2g_g_v1.0.py previous code for model training, use matlab data file as input.
  • The .mat file should contain two fields, the same as instructed above in the generate-dataset.py

$ python i2g_g_v1.0.py <data.mat (v7.3)> | tee <logfile>.txt

• i2g_train.py train the model with .npz dataset generated by generate-dataset.py, save model file .h5., it can be trained from a pre-trained model

$ python i2g_train.py <data.npz> <save.h5> [<pre-trained model.h5>]

• train_whole_and_face_roi.py train the model with .npz dataset generated by generate-dataset-v5.0.py, save model file .h5., it can be trained from a pre-trained model

$ python train_whole_and_face_roi.py <data.npz> <save.h5> [<pre-trained model.h5>]

Model prediction and visualizing the results

• predict-gaze.py predict gaze point and display both the orginal image and gaze point with trained model .h5 and file list .txt.
  • The setting is the same as the MPIIFaceGaze data generation

$ python predict-gaze.py <model.h5> <filelist.txt (e.g., p00.txt)>

Standard

• Python code should follow Google style