Mg-Gaze: multilevel gaze-aware network for appearance-based gaze estimation

Abstract

Appearance-based gaze estimation faces several challenges in complex scenarios, including the loss of local eye details, insufficient use of multi-level visual cues, and limited task adaptability of general pre-trained features. To address these problems, this paper proposes MG-Gaze, a multi-level gaze-aware network for appearance-based gaze estimation. Built upon an MAE-pretrained ViT, MG-Gaze transfers large-scale self-supervised visual representations to the gaze estimation task and designs a Multi-level Feature Fusion Module to adaptively integrate local texture, facial structure, and pose cues from different encoding levels. Furthermore, the proposed Adaptive Token Enhancement Unit recalibrates the fused tokens in a task-relevant manner and suppresses irrelevant information. Learnable gaze queries are then used to actively aggregate key regional information related to gaze direction, thereby constructing more discriminative gaze features. On MPIIFaceGaze, Gaze360, ETH-XGaze, and EYEDIAP, the proposed method achieves mean angular errors of 3.90 degree, 9.61 degree, 5.20 degree and 4.17 degree, respectively, demonstrating its clear advantage for gaze estimation.