SEEN: A Convolutional Spiking Neural Network for Efficient Pupil Coordinate Prediction from Event Data

Near-eye pupil tracking is essential for Virtual and Augmented Reality applications but poses challenges in resource-constrained environments due to the high computational demands of traditional frame-based systems. This work introduces SEEN, a lightweight Spiking Neural Network (SNN) with 144,687 parameters, designed for efficient eye tracking using Event-based Camera (EbC) data. Leveraging the sparse, asynchronous nature of EbCs and SNNs, SEEN processes Time-Surface representations of eye movements (random, saccades, reading, smooth pursuit, blinks) from the 3ET+ dataset, achieving an average Euclidean distance of 8.58 pixels on a subset and 7.98 pixels on the Kaggle 2025 challenge. Ablation studies reveal that, for recurrent leaky layers, applying a learnable β in upper layers closer to the input outperforms deeper placements, enhancing prediction accuracy and guiding efficient SNN design. The main contributions of this work include a SNN architecture and insights into design optimization, paving the way for future exploration of input features and feature extraction strategies in spiking neural networks for real-time eye tracking.