Data Selection for Reduced Training Effort in Vandalism Sound Event Detection

Typical sound event detection (SED) applications, employed in
real environments, generate huge amounts of unlabeled data
each day. These data can potentially be used to re-train the
underlying machine learning models. However, as the labeling
budget is usually restricted, active learning plays a vital role in
re-training. Especially for applications with sparse event occurrence,
a data selection process is paramount. In this paper we (i)
introduce a novel application for vandalism SED, and (ii) analyze
an active learning scheme for reduced training and annotation
effort.
In the presented system, the employed machine learning classifier
shall recognize various acts of vandalism, i.e., glass breakage
and graffiti spraying. To this end, we utilize embeddings generated
with a pre-trained network and train a recurrent neural
network for event detection. The applied data selection strategy
is based on a mismatch-first, farthest-traversal approach and is
compared to an upper bound by using all available data. Furthermore,
results for the active learning scheme are evaluated
with respect to different labeling budgets and compared to an
active learning scheme with a random sampling scheme.