Fall-Detection Model’s 94% Accuracy Was a Mirage; Leak Cut It to 69%
The author discovered that random frame-level train_test_split caused data leakage, inflating accuracy from 69% to 94% on a fall detection system using MediaPipe Pose and Random Forest. By grouping splits per video, the honest performance dropped 25 points, highlighting a common pitfall in sequential data pipelines and the need for proper evaluation design.