Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity

Summary

Real-world falls are extremely rare: the authors estimate that collecting 100 of them requires roughly 100,000 days of monitoring. That scarcity pushes the field toward simulated fall datasets, which tend to produce impressive laboratory numbers and disappointing real-world generalization. This paper evaluates motion representations for wearable fall detection specifically under that constraint.