Efficient Appearance-Based 3D Gaze Estimation Via Structural Reparameterization

Recent advances in appearance-based gaze estimation have achieved impressive accuracy, but often at the cost of computational complexity, relying on heavy deep learning backbones that limit deployment on standard consumer hardware. To bridge this gap, this work investigates the viability of modern efficient neural networks as drop-in replacements for these heavy feature extractors. A comprehensive empirical study is conducted across diverse architectural paradigms—including CNNs, Transformers, and hybrid models—to determine if lightweight networks can sustain the representational capacity required for 3D gaze regression. The findings dispel the assumption that heavy complexity is a prerequisite for high accuracy by demonstrating that compact models can effectively match the performance of established heavy baselines. Among the evaluated architectures, structural re-parameterization is identified as the optimal approach for CPU-based inference. The final proposed system, utilizing the MobileOne backbone, achieves an angular error of 10.54^∘ on the unconstrained Gaze360 dataset, comparable to the heavy ResNet-50 baseline, while delivering real-time performance on commodity laptops. These results validate the feasibility of high-performance gaze tracking on resource-constrained devices, facilitating broader accessibility for real-world applications.