Scale-Gest cuts gesture detection energy 4x with runtime-adaptive AI

Scale-Gest is a novel runtime-adaptive framework for on-device gesture detection built by researchers Abdul Basit, Saim Rehman, and Muhammad Shafique. Instead of using a single fixed detector like typical EdgeAI deployments, Scale-Gest expands the model space into a dense family of tiny-YOLO architectures. It introduces device-calibrated ACE (Accuracy-Complexity-Energy) profiles by analyzing different model-resolution-stride operating points. A lightweight runtime controller selects the appropriate ACE mode based on user-defined constraints and battery levels, while a motion-aware hand-gesture-tracking ROI gate crops the input to reduce complexity.

On a battery-powered laptop processing gesture streams, the ACE controller reduces per-frame energy by 4× (from 6.9 to 1.6 mJ) while maintaining high gesture-detection performance with event-level F1 scores of 0.8–0.9 and low mean latency of 6 ms. The team also introduces a temporally-annotated Driver Simulated Gesture (DSG-18) dataset for evaluation in real-world car driving scenarios. Accepted at DAC 2026, Scale-Gest demonstrates that runtime model selection can significantly improve energy efficiency without sacrificing accuracy—critical for mobile and embedded devices where battery life is a primary constraint.