JultikaUniversity of Oulu repository

Abstract

Successful machine learning and computer vision approach generally require significant amounts of annotated data for learning. These methods including identification, retrieval, classification of events, and analysis of human behavior from a video. Micro-level human behavior analysis usually requires laborious efforts for obtaining the precise labels. As the quantity of online video grows, the crowdsourcing approach provides a method for workers without a professional background to complete the annotation task. These workers require training to understand implicit knowledge of human behavior. The motivation of this study was to enhance the interaction between annotation workers for training purposes. By observing experienced local researchers in Oulu, the key problem with annotation is the precision of the results. The goal of this study was to provide training tools for people to improve the label quality, it illustrates the importance of training. In this study, a new annotation tool was developed to test workers’ performance in reviewing other annotations. This tool filters very noisy input by comment and vote feature. The result indicated that users were more likely to annotate micro behavior and time that refer to other opinions, and it was a more effective and reliable way to train. Besides, this study reported the development process with React and Firebase, it emphasized the use of more Web resources and tools to develop annotation tools.