Evaluating Sources of Diversity in Text-Driven IMU Synthesis for Wearable Human Activity Recognition

Lars O. Haeusler, Lena Uhlenberg, Oliver Amft (University of Freiburg, Intelligent Embedded Systems Lab)

Summary

Freiburg (Amft lab) systematically stress-tests Text2IMU: text → motion synthesis → body surface models → virtual wearable ACC/gyro, then trains HAR on synthetic-only data. They ablate diversity sources — number of textual prompts, surface-model morphologies, virtual sensor locations — and compare / combine motion-synthesis backbones. Headline: for correctly synthesised activities, best config reaches 89.2% balanced accuracy under synthetic-only training.