Reproduce “Beyond Physical Labels: Redefining Domains for Robust WiFi-based Gesture Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiang Zhang, Huan Yan, Jinyang Huang, Bin Liu, Yuanhao Feng, Jianchun Liu, Meng Li, Fusang Zhang and 1 more
GesFi redefines domains for WiFi-based gesture recognition by mining statistically coherent groups from CSI features instead of assigning domains directly from location, orientation, user, or environment labels. Across Widar3, ARIL, XRF55, and a real-world dataset, the approach improves many cross-domain settings, reaching up to 99.26% cross-location accuracy on Widar3 and 39.67% cross-dataset accuracy, while showing that physical labels are unreliable proxies for distribution shifts.
Key insights
- Physical labels do not consistently provide useful adversarial supervision: labels tied to the actual test variation can underperform unrelated labels or even a non-adversarial baseline.
- Useful latent domains mix multiple physical factors rather than corresponding cleanly to individual users, locations, or orientations, suggesting that WiFi distribution shifts are compound and nonlinear.
- Combining phase with Doppler-frequency information improves recognition, whereas adding amplitude can hurt because it is more affected by static reflections; domain weighting and gesture-adversarial clustering also contribute to performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.