Comparing Three On-Device Classifier Deployments for Ring Microgesture Recognition

Abstract

Smart rings are becoming a common wearable platform, making low-power on-device gesture recognition increasingly important. Recognizers based on inertial measurement units (IMUs) can place classification inside the sensor or on the microcontroller, but their system-level trade-offs remain unclear. We compare three on-device deployments on one custom ring: an in-sensor decision tree executed by the IMU's machine-learning core (MLC), an MCU-side convolutional neural network (CNN), and an MCU-side hyperdimensional computing (HDC) classifier. The CNN achieves the highest within-user and mean leave-one-participant-out (LOPO) macro-F1, but the MLC consumes the least measured system power (6.64 mW) and produces roughly a quarter as many false activations as the CNN on a free-living recording. Raw-sample HDC trails; an offline ablation with window-level statistical features raises its within-user macro-F1 from 0.691 to 0.904, indicating that representation contributes substantially to its window-level gap. For this platform and task, the MLC provides the most resource-efficient operating point, whereas the CNN is the accuracy-first choice.