/PoseFall

Siyuan Peng, Kate Ladenheim, Snehesh Shrestha, Cornelia Fermüller

Primary LanguagePython

Teasing Image

Introduction :

In this project, we take advantage of Captury MoCap system to capture a series of nuanced, artistically inspired falling movements. These movements serve as the foundation for training our machine learning model, the 'Attribute-Conditioned Variational Autoencoder'. With this model, we're able to synthesize a vast collection of falling motions that are not only realistic and diverse but also infused with artistic expression. To facilitate deeper exploration and interaction by artists, we've developed a comprehensive interactive interface. This tool enables users to observe the generated motions from multiple angles, control playback seamlessly, and apply the motions to a variety of human avatars. Our aim is to open up new possibilities in digital artistry, offering a bridge between technological innovation and the rich domain of human movement.

Falling Attributes:

Impact Phase

  • Impact Location: Heads, Torso, Arms, Legs
  • Impact Attribute: Push, Prick, Shot, Contraction, Explosion

Glitch Phase

  • Glitch Attribute: Shake, Flail, Flash, Stutter, Contort, Stumble, Spin
  • Glitch Speed: Slow, Medium, Fast

Falling Phase:

  • Falling Attribute: Release, Let go, Hinge, Surrender
  • End Position: Back, Front
  • End Shape: Extended, Crumpled

attribute-conditioned variational autoencoder

model_arch

Interactive Demo

<iframe src='/PoseFallWeb/' style='width: 100%;height: 40rem;'></iframe>

Quanlitative Result

result_1 result_2 result_3

note: the interactive demo is only for demo purpose, as it only load one set of seqeunce.