Reproduce “FrameScout: Scouting Query-Relevant Frames for Long Video Understanding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haonan Hu, Shuhao Chen, Weisen Jiang, Lizhao Gao, James Kwok, Yu Zhang
FrameScout selects question-relevant frames from long videos before a video question-answering model answers. It contextualizes frames across chunks and trains relevance matching with teacher-attention pseudo-labels. On VideoMME’s Long subset, it reaches 61.2% accuracy with Qwen3-VL-8B, versus 54.7% for uniform sampling using 60 frames. For eight queries, it is 6.3× faster than FlexSelect; results are limited to the tested benchmarks and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.