Reproduce “Continuous-Time Human Motion Field from Events” 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
Ziyun Wang, Ruijun Zhang, Zi-Yan Liu, Yufu Wang, Kostas Daniilidis
EvHuman estimates a continuous-time human motion field directly from event camera data, enabling efficient, parallel queries of human pose at arbitrary temporal resolutions. The method achieved improved accuracy and reduced computational time for high-speed human motion recovery when compared to prior discrete-time approaches.
Key insights
- Representing human motion as a continuous-time field via a time-implicit function (MLP) from learned latent codes enables efficient, parallel pose querying at arbitrary temporal resolutions, overcoming discrete-time limitations.
- A self-supervisory event contrast maximization loss, derived from explicit vertex optical flow, effectively trains the model using only event data for fine-grained human mesh motion, mitigating the need for dense ground truth.
- High-fidelity, hardware-synchronized datasets with agile motions, such as the newly introduced BEAHM, are essential for benchmarking and advancing continuous-time, high-speed human motion estimation from event cameras.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.