Lightweight Gaze Tracking Model Based on GhostNet
Jing-Ming Guo, Yu-Sung Cheng, Yi‐Chong Zeng, Zong-Yan Yang · 2025
In previous lightweight methods, most efforts focused on reducing the number of model parameters to lower computational requirements for terminal deployment. The aim of minimizing parameters is to achieve faster inference speeds. However, this often leads to a significant drop in accuracy. This paper seeks to accelerate model inference speed while maintaining accuracy with fewer parameters, facilitating easier deployment on terminal equipment. The proposed method employs the lightweight model GhostNet with a specialized-designed depthwise convolution. We further introduce a novel block to enhance the L2CS gaze tracking model. This approach allows the gaze tracking model to utilize a lightweight architecture from the training phase to inference one, achieving real-time performance while maintaining acceptable accuracy. Experiment results demonstrate that, compared to the original model, our improved model reduces computational parameters and inference time by nearly half, with only a 1.5% decrease in accuracy.