アイテム詳細

要旨

Abstract. Children’s social and physical environment plays a signifi-
cant role in their cognitive development. Therefore, children’s lived expe-
riences are important to developmental psychologists. The traditional
way of studying everyday experiences has become a bottleneck because
it relies on short recordings and manual coding. Designing a non-invasive
child-friendly recording setup and automating the coding process can
potentially improve the research standards by allowing researchers to
study longer and more diverse aspects of experience. We leverage mod-
ern computer vision tools and techniques to address this problem. We
present a simple and non-invasive video recording setup and collect ego-
centric data from children. We test the state-of-the-art object detectors
and observe that egocentric videos from children are a challenging prob-
lem, indicated by the low mean Average Precision of state-of-the-art.
The performance of these object detectors can be improved through fine-
tuning. Once accurate object detection has been achieved, other ques-
tions, such as human-object interaction and scene understanding, can
be answered. Developing an automatic processing pipeline may provide
an important tool for developmental psychologists to study variation in
everyday experience.