✅   Video DataFlywheel: Resolving the Impossible Data Trinity in Video-Language Understanding

science

Recently, video-language understanding has achieved great success through large-scale pre-training. However, data scarcity remains a prevailing challenge. This study quantitatively reveals an "impossible trinity" among data quantity, diversity, and quality in pre-training datasets. Recent efforts seek to refine large-scale, diverse ASR datasets compromised by low quality through synthetic annotations. These methods successfully refine the original annotations by leveraging useful information in...

NCBI     Mar 03, 2025