Reproduce “EgoWitness: Benchmarking Instance-Level Spatial Memory and Reasoning in Egocentric Video Streams” 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
Zhengru Fang, Zehao Liu, Mingyu Tsoi, Yaoyu Zhu, Junlong Li, Fengmao Lv, Sijie Cheng, Yuanliang Sun and 2 more
EgoClue uses a trained controller and visual and retrieval tools to answer EgoWitness questions that link sightings to physical objects, wearers, times, and rooms. On the same 1,320-question week-long evaluation, Micro accuracy rises from 23.64% for the prompted controller to 72.65% after SFT and GRPO. On 787 cross-environment questions spanning four abilities, it reaches 78.40%, not general performance across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.