Answers are grounded in this paper and cite their sources.
Eye tracking is a crucial task for smart eyewear; however, conventional frame-based pipelines are costly for always-on settings because they rely on dense image acquisition and processing. Event cameras and spiking neural networks (SNNs) provide a promising alternative; however, practical embedded eye tracking remains limited by the difficulty of regressing continuous pupil coordinates from sparse spikes and the gap between off-chip training and neuromorphic deployment. This paper presents SpELL, fully Spiking Eye-tracking at Low Latency, a fully spiking event-based eye-tracking framework for neuromorphic hardware. The proposed system combines a fully hardware-compatible SNN for direct pupil-center estimation and a preprocessing stage that controls event throughput. The method is evaluated on the Angelopoulos, Ini-30, and 3ET+ datasets and deployed on the SPECK neuromorphic processor. Off-chip, SpELL achieves a mean error of 2.83 px across the three datasets. On-chip, SpELL achieves competitive tracking accuracy with SPECK chip power consumption as low as 2 mW, excluding the external processing used for input regulation and output readout. These results demonstrate the potential of fully spiking inference for low-power eye tracking in wearable event-based vision systems.
No takes yet. Share an insight, caveat, or question.
Simpsi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: