Reducing Label Fragmentation During Time-series Data Annotation to Reduce Annotation Costs
Joseph Korpela, Takayuki Akiyama, Takehiro Niikura, Katsuyuki Nakamura · 2021
Labeling training data for human activity recognition systems is a time consuming task that is prone to errors. In this paper, we build on prior works that have proposed semi-automatic labeling techniques to improve the labeling process by proposing a novel technique for reducing label fragmentation in time-series data that can reduce annotation costs by improving how well the automatically generated labels model the time intervals of the underlying activities. We perform temporal clustering of the classifier output using an image segmentation algorithm, with the time-series output of the classifier fed to the algorithm as a 1D image with the class probabilities in place of color channels. During an evaluation of the proposed method on the task of labeling 233 minutes of time-series data, we achieved a 56% average reduction in annotation time when compared to the use of raw classifier output for semi-automatic labeling.