OA18: A New Office Actions Benchmark

Abstract

Action recognition is one of the current hot topics in the literature due to its important applications. To better evaluate the existing models, several datasets are publicly available and range in scale from synthetic datasets to complex realistic datasets. However, it seems that very few datasets are domain specific. This paper publishes a new recorded dataset that studies 18 domain specific human actions, specifically, actions performed by employees inside an office. This is the first published dataset in the literature for the office domain. Moreover, a python-based software is developed and used for labeling the recoded videos and presented for future usages. The recorded dataset is designed not only to study several existing challenges in the literature which include the variation of viewpoints for the same class and the issue of shaky videos, but we also cover the usage of deep learning techniques on this small dataset by performing data augmentation and transfer learning.