ExpressEar: Sensing Fine-Grained Facial Expressions with Earables
summary
Ear-mounted 6-DOF IMU sensing on Nokia eSense captures ear-canal deformations sampled at 50 Hz, windowed into 2s segments, bandpass filtered and fed to a Temporal Convolutional Network. A dataset of 12 users yields ~120,000 labeled segments and per-user classification of 32 AUs at 89.9% per-user accuracy (leave-one-user-out 42.1%). Single-ear models hit ~79% and real-world tests show ≈84% accuracy during walking and metro travel, demonstrating a working, privacy-preserving alternative to camera FER.
AI-written summary · about