Doctoral Seminar (ICS, HCR) on "Learning how to build a vector embedding to represent low-level human actions - attempts for hierarchically understanding high-level human actions" by Dr. Hyemin Ahn/HCR
Abstract: In this seminar, the presenter will talk about her recent studies related to the human behavior understanding based on the video. To be more specific, how to hierarchically understand the observed human behavior will be discussed. The biggest assumption of this study is that in order to understand a high-level human action, it is necessary to understand the low-level human actions first that constitute the high-level human action. For example, suppose a human is making a salad (high-level action). To make a salad, this person will open the refrigerator, pulling out vegetables, turning on the faucet, and washing the vegetables (low-level actions). That is, a video observing a high-level human action can be segmented into low-level human actions, and the high-level human action can be understood as a sequence of corresponding action segments.
Regarding this, the presenter has tried various methodologies for learning a vector embedding, that can represent each low-level action segment composing the high-level human action. Rather than directly understanding a high-level human action based on the hundreds of image frames, methodologies suggested by the presenter divide the high-level human action into low-level action segments, find a set of vector embeddings for those action segments, and understand the high-level human action based on the sequence of those vector embeddings. It would be a pleasure to discuss about pros and cons of each approach, analyze why some attempts have failed or succeeded, and share the feedback with each other in an open atmosphere.
June 8th, 2020 at 11h (online). More details here.