Label-Free Adaptation for Large-Scale Aerial Crowd Monitoring

A new deployment protocol for drone-based crowd counting uses label-free adaptation to recover 31-49% of error caused by distribution shifts. The study establishes a severity law and stability budget to ensure safety in high-stakes environments like mass gatherings.