Metrological Reformation for Continuous Hand Gesture Recognition: Framework, Benchmarks, and Validation

Continuous hand gesture recognition (CHGR) suffers from a critical gap between benchmark performance and real-world deployment, driven by flawed evaluation metrics and poor non-gesture modeling. This work addresses these issues with three novel technical contributions and two new benchmarks for CHGR: (1) a formal analysis of frame-based metric bias in CHGR, showing rank inversion and 47 percentage point (pp) duration-dependent variation in Jaccard Index (JI); (2) GeNGA , a diffusion-based generative non-gesture augmentation framework that reduces false positive rates (FPR) by 47.3% on naturalistic data; (3) the Dictionary Difficulty Index (DDI) , a quantitative metric explaining 73% of cross-benchmark CHGR performance variance; and (4) the Continuous Gesture Metrology Framework (CGMF) , a standardized evaluation protocol for CHGR. We also release two CGMF-compliant benchmarks: MIX-HAND200 (207 subjects, 100h non-gesture data) and AutoGest-Drive (60 drivers, 45h automotive gesticulation). Empirical re-evaluation of 3 state-of-the-art (SOTA) CHGR methods on 4 benchmarks shows reported JI overstates true event detection rate (EDR) by 9–20.5 pp, with SOTA EDR at 60–73% under validated thresholds. GeNGA and CGMF together reduce real-world FPR by 60.8%, demonstrating transformative improvements for CHGR deployment.