Effective Performance Metrics for Evaluating Activity Recognition Methods

Inhalt:
In this paper, we present several novel metrics for evaluating the recognition performance of activity recognition methods. Traditional methods of evaluation are insufficient for activity recognition because they do not take into account class imbalance and do not account for errors specific to temporal data problems. We present several metrics that do take these issues into account. The effectiveness of our approach is shown by comparing the recognition performance of two closely related probabilistic models on three real world datasets.