Event-Based Eye Tracking. 2025 Event-based Vision Workshop
This work addresses real-time, high-accuracy pupil center prediction from event camera–acquired eye movement data, advancing the practical deployment of event-driven eye tracking. We systematically survey and benchmark top-performing methods from the Eye Tracking Challenge at the 2025 CVPR Workshop on Event-Based Vision. Our proposed solution is the first to jointly integrate spiking neural networks, spatiotemporal event stream encoding, a lightweight CNN-LSTM hybrid architecture, and low-latency online filtering—augmented by hardware-aware co-design principles. The resulting optimal model achieves Pareto-optimal trade-offs among accuracy (mean error < 2.1 pixels), model size (< 300K parameters), and computational cost (< 1.2 GOPS), significantly outperforming conventional frame-based approaches. This end-to-end solution enables low-power, high-responsiveness eye tracking with direct deployability on resource-constrained edge platforms.