Fig. 2

Structure of the proposed EM-Gaze network. Given an input face image, facial features are extracted by Label-Net and Face-Net. Eye-Net takes the left eye image and flips right eye image as inputs, extracts eye features under the guidance of facial features, and iteratively correlates two-eye features using the proposed CCBs. Concatenated eye and facial features are fed into channel-mixing layers to obtain the gaze feature. Finally, fully-connected layers are employed to estimate 2D gaze position and quadrant division-based classification