Fusing video and text data by integrating appearance and behavior similarity

摘要

In this paper, we describe an algorithm for multi-modal entity co-reference resolution and present experimental results using text and motion imagery data sources. Our model generates probabilistic association between entities mentioned in text and detected in video data by jointly optimizing the measure of appearance and behavior similarity. Appearance similarity is calculated as a match between proposition-derived entity attributes mentioned in text, and the object appearance classification from video sources. The behavior similarity is calculated based on the semantic information about entity movements, actions, and interactions with other entities mentioned in text and detected in video sources. Our model achieved 79% F-score for text-to-video entity co-reference resolution; we show that entity interactions present unique features for resolving variability present in text data and ambiguity of visual appearance of entities.