A new benchmark gives AI models a seemingly simple task: find specific photos in a personal collection.
When people look for a specific photo, they usually remember the context rather than the image itself. The concert photo where only the singer was visible, from the show with the blue and white logo at the entrance.
The key clue to which concert that actually was is hidden in a completely different image. According to a new study by researchers at Renmin University of China and the research institute of smartphone manufacturer Oppo, this is exactly where every standard image search system falls apart.
Today’s multimodal search systems evaluate each image on its own: does it match the query or not? That works fine when the target photo is visually distinctive. But as…