Official code for "Revisiting Subject-Independent Evaluation for Smartphone-Based Fall and Activity Recognition". This repository quantifies the CV-to-LOSO generalization gap on the MobiAct dataset and introduces Feature-Invariance-Conditioned (FIC) pooling with SAM to improve cross-subject HAR performance.
human-activity-recognition attention-mechanism fic sharpness-aware-minimization loso cross-subject-generalization mobiact
-
Updated
Aug 26, 2026 - Python