CI-Net: Appearance-Based Gaze Estimation via Cooperative Network

Abstract

Facial occlusion and different appearances of both eyes in natural scenes can affect the accuracy of gaze estimation based on appearance. Therefore, this paper proposes a gaze estimation model based on cooperative network: CI-Net, including a consistency estimation network (C-Net) and inconsistency estimation network (I-Net). C-Net is used to estimate the Main gaze of the true gaze, and an attention mechanism is added to adaptively assign the weight between eyes and face features. The I-Net is used to estimate the Residual residuals based on true gaze. In addition, Cross attention module is designed in this paper, through which I-Net can selectively obtain information from C-Net, to obtain more accurate eyes directions. The experimental results in this paper show that the CI-Net gain lower angle errors than the current mainstream CNN methods under the condition of different appearance of both eyes and facial occlusion.