[1H5-GS-10] Vision, speech: Recognition and detection

[1H5-GS-10-05] A Video Dataset for Action Detection and Understanding in Nursery Schools

〇Keita Iida¹, Xueting Wang¹, Toshihiko Yamasaki¹, Satoshi Toriumi², Mikihisa Hayashi², Sachiko Nozawa¹, Midori Takahashi¹, Kengo Hiroto¹, Toshihiko Endou¹, Kiyomi Akita¹ (1. The University of Tokyo, 2. Future Standard Co., Ltd.)

Keywords:Action detection, Actio understanding, Dataset

Action detection, which is a task whose goal is to find person (action localization) and classify what he/she is doing (action classification), is a challenging task in computer vision. Despite its difficulty, various methods for action detection from videos have been developed recently. However, there are not many datasets for detailed action understanding because of great expense of dataset construction.
We introduce a new annotated video dataset for action understanding, based on videos taken in nursery schools. In this dataset, we give detailed tags for every person in the videos, which enable advanced analysis in actions of children. We focus on interactions with their surroundings, by annotating not only their actions but also their IDs and objectives of the actions.