WHAR Arena: Benchmarking the State of the Art in Efficient Wearable Human Activity Recognition
A new benchmark called WHAR Arena has been developed to address the comparability crisis in Wearable Human Activity Recognition (WHAR) deep learning research. This open-source benchmark standardizes datasets, processing, and evaluation protocols across 30 datasets and 17 architectures. The findings indicate that while predictive performance has plateaued, there is significant potential for progress in optimizing deployment efficiency, particularly for compact models like TinierHAR and classical Random Forests, which offer a better balance of performance and hardware cost compared to larger recurrent and hybrid models. AI
IMPACT Highlights the trade-offs between predictive performance and deployment efficiency in wearable AI, guiding future research towards practical applications.