Domain Generalized Gaze Estimation Network for Enhanced Cross-Environment Performance

Gaze estimation technology provides a non-invasive and natural interaction method by tracking the direction of a person's gaze, making it valuable for various applications. However, existing gaze estimation models are often optimized for specific environments, and their performance tends to degrade significantly when applied to new or different settings. This decline in performance is primarily due to substantial variations in factors such as lighting conditions, ethnicity, gender, and age between different environments, resulting in insufficient generalization capabilities of the models. To address the aforementioned challenges, we propose a novel gaze estimation network, GE-Net, which significantly enhances the cross-environment performance of eye-tracking models through domain generalization. The network integrates Frequency Domain Data Augmentation (FDDA) with a lightweight Multilayer Perceptron (LightMLP), allowing it to learn gaze-consistent features without the need for target domain data sampling. The architecture of GE-Net comprises a feature extraction backbone (ResNet18), an image restoration module, and the innovative LightMLP. By perturbing gaze-irrelevant features in the frequency domain, the model demonstrates improved generalization capabilities across different datasets. Experimental results on four cross-domain datasets demonstrate that GE-Net exhibits superior generalization capabilities compared to existing methods, even in the absence of target domain samples. The network's efficiency and adaptability make it a promising tool for applications requiring reliable gaze estimation. The implementation code can be found at https://github.com/yuhanghong123/GENet.