Automatic correction of annotation boundaries in activity datasets by class separation maximization

Abstract

t is challenging to precisely identify the boundary of activities in order to annotate the activity datasets required to train activity recognition systems. This is the case for experts, as well as non-experts who may be recruited for crowd-sourcing paradigms to reduce the annotation effort or speed up the process by distributing the task over multiple annotators. We present a method to automatically adjust annotation boundaries, presuming a correct annotation label, but imprecise boundaries, otherwise known as "label jitter". The approach maximizes the Fukunaga Class-Separability, applied to time series. Evaluations on a standard benchmark dataset showed statistically significant improvements from the initial jittery annotations.