Event cameras provide high temporal resolution and low latency, making them attractive for gaze estimation on resource-constrained wearables. However, widespread pipelines remain sensitive to noisy 2D pupil observations, while high-accuracy learning approaches can be too heavy for always-on embedded deployment. We present an event-frame adaptation of a classical geometric pipeline based on Pupil Labs’ pye3d and a set of lightweight neural models that map compact geometric features (pupil ellipse parameters and, when available, corneal reflections) to gaze direction and 3D eye center. Across two synthetic datasets and a real dataset acquired on the Ameca humanoid robot, the best temporal model (Encoder–LSTM–Decoder) achieves mean gaze errors of 0.204◦ (Synthetic), 1.299◦ (Synthetic with Slippage, simulating smooth eye-center shifts), and 0.242◦ (Real).
Lightweight Neural Networks for Event-Based 3D Gaze Estimation on Wearable Devices
Tognoli, Aaron;Simpsi, Andrea;Aspesi, Andrea;Matteucci, Matteo;Mentasti, Simone;Cannici, Marco
2026-01-01
Abstract
Event cameras provide high temporal resolution and low latency, making them attractive for gaze estimation on resource-constrained wearables. However, widespread pipelines remain sensitive to noisy 2D pupil observations, while high-accuracy learning approaches can be too heavy for always-on embedded deployment. We present an event-frame adaptation of a classical geometric pipeline based on Pupil Labs’ pye3d and a set of lightweight neural models that map compact geometric features (pupil ellipse parameters and, when available, corneal reflections) to gaze direction and 3D eye center. Across two synthetic datasets and a real dataset acquired on the Ameca humanoid robot, the best temporal model (Encoder–LSTM–Decoder) achieves mean gaze errors of 0.204◦ (Synthetic), 1.299◦ (Synthetic with Slippage, simulating smooth eye-center shifts), and 0.242◦ (Real).I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
