Spatial Transformer Networks for Sensor-based Human Activity Recognition
Overview
Human Activity Recognition (HAR) is the task of classifying the activity of humans from wearable sensors like smartwatches. These are sometimes misplaced, e.g. worn on the wrong arm or rotated, compared to the training data. Conventionally, classifiers will not generalize to such misplaced data. In this Bachelor or Master thesis, we will explore Spatial Transformer Networks (STNs), which should be invariant to such misplacement. This work builds on recent work in STns and will be jointly supervised with the authors of that paper.
References
Johann Schmidt, Tom Siegl, Martin Becker, and Sebastian Stober. “Geometrically Constrained and Token-Based Probabilistic Spatial Transformers”. ECCV Workshops, 2026.
Contact
Stefan Lüdtke