Motion-DVAE: Unsupervised learning for fast human motion denoising

What is it about?

This work introduces Motion-DVAE, a human motion prior for modeling short-term dependencies of human movement. Together with Motion-DVAE, we propose an unsupervised learned denoising method allowing for robust noise modeling. This procedure can be used for motion denoising in regression- or optimization-based frameworks and adapts to new data in real time. Visual results are available on the Motion-DVAE webpage: https://g-fiche.github.io/research-pages/motiondvae/

Featured Image

Photo by Uriel SC on Unsplash

Photo by Uriel SC on Unsplash

Why is it important?

In summary, our contributions are: (i) A generative motion prior, Motion-DVAE, representing short-term dependencies of human motion. (ii) A flexible unsupervised learning framework for real-time human motion denoising in regression- and/or optimization-based procedures. (iii) The proposed method is faster than state-of-the-art methods while showing competitive performance on pose estimation and motion denoising.