Reproduce “A Principled Optimal Transport Framework for Frame Selection in 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
Ziwei Zhao, Yuhang Liu
OTFS allocates evidence across candidate video frames, then selects a fixed-size subset for long-video question answering. Semi-unbalanced optimal transport converts multiple evidence sources into frame weights; an exact dynamic program balances those weights with temporal spread. On MLVU, Qwen3-VL-4B achieves 72.10% accuracy with 32 frames, versus 68.27% for AKS; this measured result does not guarantee optimal downstream answers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.