Reproduce “A Short Note on the Kinetics-700-2020 Human Action Dataset” 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
Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, Andrew Zisserman
Google researchers updated the Kinetics-700 human action dataset, creating Kinetics-700-2020, to address video clip disappearance and ensure a minimum of 700 examples per class. The refreshed dataset enabled a standard I3D model to achieve improved action recognition accuracy compared to the previous version.
Key insights
- Proactive dataset maintenance, including replenishment and rebalancing, is crucial for preserving the integrity and long-term utility of large-scale, web-sourced datasets.
- Increased quantity and improved balance of training examples directly contribute to enhanced performance in action recognition models.
- Multilingual query expansion and robust duplicate detection mechanisms are effective strategies for curating higher quality and more diverse video datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.