Oxford University : Magic Pony

by Pierre Colle

Researchers from the Oxford University are able to model articulated horses, using only one viewpoint image per horse.

Main takeover from this article is the new “Implicit-Explicit” approach

“Implicit-Explicit” is mixing :

  • Explicit: Meshes are usually used to model articulated things and are trained with few images
  • Implicit: SDF (signed distance function) are used to have a fine-grained 3d models, used when having multiple view-point of the same static scene

🦵 The 3D shape of horse with legs … is difficult to model for AI, as we can see it with the 5-legged-horse given by diffusion models

By mixing both, they are able to have a complex and fine-grained model of horses in a self-supervised manner.