Violin Video and Language Inference Dataset

Violin stands for VIdeO-and-Language INference, which can be used for multimodal understanding tasks of videos and texts.

The dataset contains 95,322 video-hypothesis pairs from 15,887 video clips, covering more than 582 hours of video. These video clips contain rich content with different temporal dynamics, event changes, and human interactions. The data is collected from two sources: (i) popular TV shows, and (ii) movie clips from YouTube channels.