Retrieval-based Annotation of Multi-channel Time-Series Data for HAR
Erik Altermann, Fernando Moya Rueda, Eugen Rusakov, Gernot A. Fink · 2022 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops) · 2022
This contribution proposes a semi-automated annotation procedure for multi-channel time-series data for human activities. This procedure leverages the human annotators’ decision by employing ranked hypotheses from a deep architecture, decreasing annotation time and maintaining consistency. In general, our approach performs retrieval by querying activity classes, Query-by-Activity. First, retrieved segments are ranked w.r.t a query using a similarity metric. Then, a human annotator can accept the ranked segments instead of selecting from a set of classes and attributes for each segment. We evaluate the approach in two parts, a retrieval and an annotation task. The retrieval performance is examined via the average precision of retrieved activities on a testing set. The effort is measured in terms of annotation time and consistency when annotating a small dataset for human activities.