What does "InfiniBench" mean?
Table of Contents
InfiniBench is a new way to test how well machines can understand long videos. While most existing tests look at short clips, InfiniBench focuses on videos that can last anywhere from tens of minutes to several hours. Think of it as a marathon for AI—only instead of running, they're trying to answer questions about what they just watched!
Key Features of InfiniBench
-
Long Video Duration: The average video in InfiniBench is about 52.59 minutes long. That's nearly as long as some people's favorite binge-watching sessions!
-
Question Variety: InfiniBench includes a whopping 108,200 question-answer pairs. These questions explore nine different skills, mixing multiple-choice and open-ended formats. So whether you're asking a straightforward question or one that requires a bit of thought, InfiniBench has you covered.
-
Real Life Content: The videos come from popular films and TV shows, making them relatable. Some questions even require viewers to think critically, such as those sneaky "Movie Spoiler Questions." It's like a trivia night at home but for machines!
-
Evaluation of AI Models: InfiniBench tests various large models, including some of the leading names in AI. The results show that even the best machines struggle with long video understanding. For instance, the top models scored around 49.16% and 42.72% on average. They might need a bit more popcorn and practice!
The Purpose of InfiniBench
The goal of InfiniBench is to push the boundaries of how machines understand video content. By focusing on long videos and real-world questions, it encourages researchers to improve AI models. So next time you wonder why your streaming service doesn't quite get your taste in films, just remember, the machines are still learning!
In a world where attention spans seem to be getting shorter, InfiniBench is here to remind everyone that sometimes, good things take time.