Reproduce “From Lifestyle Vlogs to Everyday Interactions” 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
David F. Fouhey, Wei-cheng Kuo, Alexei A. Efros, Jitendra Malik
UC Berkeley researchers developed an "implicit data gathering" methodology to create VLOG, a large-scale dataset of 114,000 diverse video clips showing everyday human-object interactions. This dataset exposes significant biases in explicitly gathered scene classification and object detection datasets, and enables benchmarking of hand-centric interaction tasks.
Key insights
- Implicit data gathering from naturally occurring video genres, like lifestyle vlogs, offers a scalable and effective alternative to traditional explicit search and staged collection methods for acquiring diverse, naturalistic interaction data.
- Widely used, explicitly gathered datasets (e.g., Places365, COCO) exhibit significant biases, causing models trained on them to generalize poorly to naturalistic, unconstrained everyday interaction data found in VLOG.
- Focusing annotations on human hands proves effective for broadly capturing and understanding diverse human-object interactions in "in-the-wild" video, as hands are central to such interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.