Optimizing Gaze Estimation With a DLA-Based Calibration Module on NVIDIA Jetson Platforms

Gaze estimation is a commonly used task in human-computer interaction, with its applications ranging from driver monitoring systems to human assistive technologies. However, achieving robust, real-time performance on edge devices remains a challenge due to limited computational resources and the need for high energy efficiency. In this work, we present a novel gaze estimation model optimized for NVIDIA Jetson platforms, leveraging the unique capabilities of Deep Learning Accelerators (DLAs) for enhanced efficiency. Our model includes a calibration module designed for DLA execution, incorporating convolutional spatial and channel attention to improve robustness after face detection stage. By strategically distributing computational tasks across GPU and DLA, our approach achieves up to an 88% increase in energy efficiency compared to a system without calibration module, making it well-suited for real-time, edge-based applications in resource-constrained environments.