Exploring the Dynamics of Visual Events in the Multi-dimensional Semantic Concept Space

Abstract

We present a system for visualizing event detection in video and revealing the algorithmic and scientific insights. Visual events are viewed as evolving temporal patterns in the semantic concept space. For video clips of different events, we present their corresponding traces in the semantic concept space as the event evolves. The presentation of the event in the concept space is scored by pre-trained models of the dynamics of each concept in the context of the event, which provides a measure of how well the given event matches the evolution pattern of the target event in the multi-dimensional concept space. Scores obtained for different videos is shown to project them into different parts of the final score space. This presentation walks the user through the entire process of concept-centered event recognition for events such as exiting car, riot, and airplane flying.