Real-time eye contact correction using neural network-based machine learning

摘要

Techniques relating to eye contact correction are discussed that provide a camera-oriented virtual user view while the user is looking at a display. Such techniques may include encoding an eye region of a source image using a pre-trained neural network to generate compressed features, applying a pre-trained classifier to the features for determining a motion vector field for the eye region, and deforming and inserting the eye region into the source image to generate a Cover image with corrected eye contact.