Reproduce “OMG-Bench: A New Challenging Benchmark for Skeleton-based Online Micro Hand Gesture Recognition” 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
Haochen Chang, Pengfei Ren, Buyuan Zhang, Da Li, Tianhao Han, Haoyang Zhang, Liang Xie, Hongbo Chen and 1 more
Researchers introduce OMG-Bench, the first large-scale benchmark for skeleton-based online micro hand gesture recognition, and propose HMATr, an end-to-end framework that integrates hierarchical memory and learnable position-aware queries. HMATr achieved an 89.2% Detection Rate on OMG-Bench with an inference time of 1.61ms, outperforming existing methods for both micro and macro gesture recognition.
Key insights
- High-quality, multi-view derived skeleton data, combined with precise semi-automatic annotation, is essential for creating effective benchmarks for subtle micro gestures.
- Hierarchical temporal memory, integrating both fine-grained frame-level and abstract window-level context, is critical for robust online recognition of rapid and continuous micro gestures.
- Position-aware queries can effectively unify detection and classification, providing a flexible mechanism to localize and categorize subtle gestures within continuous data streams without relying on explicit anchors or two-stage pipelines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.