CelebAGaze

CelebAGaze consists of 25283 high-resolution celebrity images that are collected from CelebA and the Internet. It consists of 21832 face images with eyes staring at the camera and 3451 face images with eyes staring somewhere else. All images (256 × 256) are cropped and the eye mask region by dlib is computed. Specifically, dlib is used to extract 68 facial landmarks and calculate the mean of 6 points near the eye region, which will be the center point of the mask. The size of the mask is fixed to 30×50. As described above, 300 samples from domain Y are randomly selected, 100 samples from domain X as the test set, the remaining as the training set. Note that this dataset is unpaired and it is not labeled with the specific eye angle or the head pose information.

Source: Dual In-painting Model for Unsupervised Gaze Correction and Animation in the Wild