SigGen — Controllable wireless signal generation
Generating wireless signals for unseen combinations of people and environments.
The problem
Collecting wireless sensing data across different people and environments is time-consuming and costly. Generating signals for unobserved combinations offers a way to expand training data without collecting every combination.
The approach
We learn disentangled representations that separate human-related information from environmental factors in wireless signals. Recombining these representations enables controllable synthesis—for example, using signals from person A in room 1 and person B in room 2 to generate signals for person B in room 1.
Results
The method improves signal generation quality by 30.83% over state-of-the-art methods. Using the generated signals for data augmentation also improves gesture recognition accuracy by 12.74%.