What Is the Role of Similarity for Known-Item Search at Video Browser Showdown?

摘要

Across many domains, machine learning approaches start to compete with human experts in tasks originally considered as very difficult for automation. However, effective retrieval of general video shots still represents an issue due to their variability, complexity and insufficiency of training sets. In addition, users can face problems trying to formulate their search intents in a given query interface. Hence, many systems still rely also on interactive human-machine cooperation to boost effectiveness of the retrieval process. In this paper, we present our experience with known-item search tasks in the Video Browser Showdown competition, where participating interactive video retrieval systems mostly rely on various similarity models. We discuss the observed difficulty of known-item search tasks, categorize employed interaction components (relying on similarity models) and inspect successful interactive known-item searches from the recent iteration of the competition. Finally, open similarity search challenges for known-item search in video are presented.