Abstract: With the wide application of convolutional neural network (CNN) in the field of computer vision and the release of a large number of 3D gaze datasets, research on 3D gaze estimation based on the combination of apparent and deep learning has received more and more attention. However, due to the complex structure of CNN, such methods need to be further improved in occasions with high real-time requirements. Recent studies have shown that MLP models with simpler structures can achieve performance comparable to the current best CNN and Transformer models. Inspired by this, an efficient and high-precision 3D gaze estimation method based on MLP is proposed. The MLP model is used to extract features from face and binocular images and then fuse them to derive 3D gaze. Experiment shows that, for the 31 subjects with different appearance characteristics in MPIIFaceGaze dataset and EyeDiap dataset, the proposed method UM-Net achieves gaze estimation accuracy that is comparable to CNNs-based method, and it has obvious advantages in gaze estimation speed. Therefore, it has a good application prospect in fields with high real-time requirements.

Key words: 3D gaze estimation, appearance, multi-layer perceptron(MLP), real-time