Reproduce “GigaHands: A Massive Annotated Dataset of Bimanual Hand Activities” 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
Rao Fu, Dingxi Zhang, Alex Jiang, Wanjia Fu, Austin Funk, Daniel Ritchie, Srinath Sridhar
Researchers from Brown University and ETH Zurich introduce GigaHands, a large-scale, richly annotated 3D dataset of bimanual hand activities featuring 34 hours of motion and 84,000 text descriptions. This dataset enables enhanced performance in text-driven hand motion synthesis and captioning, and supports high-fidelity dynamic radiance field reconstruction.
Key insights
- Combining procedural instruction generation with LLM augmentation significantly enhances data diversity and reduces manual annotation effort for complex bimanual activities.
- A large-scale, high-fidelity dataset with continuous 3D hand and object poses, dense multi-view imagery, and rich text annotations substantially improves performance in multimodal hand activity understanding and generation tasks.
- The 'Instruct-to-Annotate' methodology provides a balanced approach to capture accuracy, diversity, realism, and scalability, overcoming limitations of previous in-the-wild or marker-based studio setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.