Multi Scale Super Resolution Gaze Estimation Networks Embedded with Cross Attention

Although many gaze target estimation methods have been proposed, most gaze target estimation methods are not very effective in processing low-resolution images, especially low-resolution images need to be further cut the head box, in addition, the lack of local and global relationship is also a problem of current methods. In order to solve the above problems, a multi-scale super-resolution fixation estimation network embedded with cross-attention is proposed. Our network improves the resolution of each scale feature of the input image and enhances the high frequency information such as the eye feature and its boundary based on the multi-scale super resolution module to obtain more abundant facial high frequency features. In the implementation of the module, local global feature fusion is used: the eye area is segmented from the face image, and the face and eyes are respectively enhanced with high frequency, and the features are superimposed according to the position coding of the eyes. In addition, the cross-attention module is used to enhance the connection between local and global, to strengthen the local and global relationship by calculating the attention relationship between the head features and the scene features, and to use the obtained attention results to regression the gaze target. The experimental results show that the proposed method can perform robust gaze estimation even in low resolution face images.