TempCloze is a video cloze benchmark that requires models to identify the correct middle clip from a sequence to eliminate linguistic shortcuts. It uses 1,521 videos to test semantic, alignment, and progression capabilities through visual temporal reasoning.
HOW THIS AFFECTS YOU
●
researcherYou can use this to ensure your video models are actually processing temporal data rather than relying on text cues.