Inf-Stream-Eval is a benchmark for evaluating vision-language models (VLMs) on near-infinite video streams. It consists of videos averaging over two hours in length that require dense, per-second alignment between video frames and text.
This dataset was introduced in the paper StreamingVLM: Real-Time Understanding for Infinite Video Streams.
