Evaluating the transferability of personalised exercise recognition models.

Exercise Recognition (ExR) is relevant in many high impact domains, from health care to recreational activities to sports sciences. Like Human Activity Recognition (HAR), ExR faces many challenges when deployed in the real-world. For instance, typical lab performances of Machine Learning models, are hard to replicate, due to differences in personal nuances, traits and ambulatory rhythms. Thus effective transferability of a trained ExR model, depends on its ability to adapt and personalise to new users or user groups. This calls for new experimental design strategies that are also person-aware, and able to organise train and test data differently from standard ML practice. Speciffically, we look at person-agnostic and person-aware methods of train-test data creation, and compare them to identify best practices on a comparative study of personalised ExR model transfer. Our findings show that ExR when compared to results with other HAR tasks, to be a far more challenging personalisation problem and also confirms the utility of metric learning algorithms for personalised model transfer.