Exquisitor at the Video Browser Showdown 2022

Omar Shahbaz Khan, Ujjwal Sharma, Björn Þór Jónsson, Dennis C. Koelma, Stevan Rudinac, Marcel Worring, Jan Zahálka

Abstract: Exquisitor is the state-of-the-art large-scale interactive learning approach for media exploration that utilizes user relevance feedback at its core and is capable of interacting with collections containing more than 100M multimedia items at sub-second latency. In this work, we propose improvements to Exquisitor that include new features extracted at shot level for semantic concepts, scenes and actions. In addition, we introduce extensions to the video summary interface providing a better overview of the shots. Finally, we replace a simple keyword search featured in the previous versions of the system with a semantic search based on modern contextual representations.