E-SpikePT: Energy-efficient event-based eye tracking with spiking neural networks and knowledge distillation

초록

Event-based eye-tracking enables high-speed and low-power visual perception by leveraging the asynchronous nature of neuromorphic sensors. However, most existing approaches still rely on computationally intensive artificial neural networks (ANNs), which limit system-level energy efficiency during inference. In this paper, we propose E-SpikePT, an energy-efficient event-based pupil tracking framework that integrates spiking neural networks (SNNs) with knowledge distillation. Our architecture consists of a Spiking Residual Temporal Dynamics (SRTD) module for short-term spatio-temporal feature extraction and a Spiking Temporal Dynamics (STD) module that combines a GRU and Mamba-based state-space modeling with Parametric Leaky Integrate-and-Fire (PLIF) neurons. To mitigate the performance degradation commonly observed in SNNs, we introduce a coordinate-aware knowledge distillation strategy tailored for eye-tracking, transferring spatial precision from a high-capacity ANN teacher to the SNN student without additional inference overhead. Extensive experiments on two event-based eye-tracking benchmarks, SEET and 3ET+, demonstrate that E-SpikePT achieves competitive tracking accuracy (MSE of 1.56 on 3ET+) while significantly reducing power consumption to 0.049mJ. © 2026 Elsevier B.V.

키워드