Reproduce “Hollywood in Homes: Crowdsourcing Data Collection for Activity 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
Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, Abhinav Gupta
The "Hollywood in Homes" methodology collected the Charades dataset, comprising nearly 10,000 videos of diverse, everyday household activities with detailed temporal and object interaction annotations. This new resource provides realistic data for training AI models, addressing the limitations of existing datasets biased towards entertainment or controlled environments.
Key insights
- Crowdsourcing can effectively scale the creation of diverse, real-world video content, overcoming the inherent biases and limitations of purely lab-controlled or internet-sourced datasets.
- Charades videos contain a high density of activities, averaging 6.8 relevant actions per video, providing rich contextual information and sequences reflecting common sense.
- The dataset's focus on fine-grained human-object interactions in diverse household settings exposes a core challenge for current computer vision models in differentiating similar actions with the same objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.