The research team introduced SafeTransLoc, which compares a transferred model with the best model trained solely on target-building data before deployment across buildings. Transfer is permitted only when validation error meets a predefined safety margin. The experiments covered 120 cross-building configurations from UJIIndoorLoc and SODIndoorLoc, and the transferred models never outperformed the strongest target models. With a 0.5-meter safety margin, the method rejected 100% of harmful transfers on the original datasets.
Source:BIOENGINEER.ORG · Read Original
Leave a Reply