Reproduce “PhysCapBench: A Cross-Domain Benchmark for Physically Grounded Video Captioning” 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

Haofeng Wang, Xuanyi Liu, Pengcheng Wen, Ruiqi Li, Yuhao Zhang, Tiange Zhang, Boren Zheng, Siyu Tao and 7 more

PhysCapBench evaluates whether video captions preserve physical events across third-person, egocentric, and handheld-gripper videos. It contains 1,500 human-verified sequences and scores captions against verified references along seven dimensions. In reconstruction tests, physical captions reduced perceptual distances across both tested generators and domains; for MiniMax H3 on third-person videos, LPIPS fell 9.51%, without establishing physical correctness or covering gripper caption strategies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.