Human Gaussian Splats – Ability to create animable (3DGS) avatars from casual videos (50-100 frames) in about 30 minutes. Creator avatars can be easily embedded in other (NeRF) scenes.
Item: https://machinelearning.apple.com/research/hugs
arXiv： https://arxiv.org/abs/2311.17910
HUGS learns the separation representation of humans (in canonical space) and scenes, allowing them to edit and animate human avatars in scenes.
HUGS is based on 3D Gaussian stitching (3DGS) and SMPL manikins to drive 3D human Gaussian; And allows the model to capture details beyond the mannequin, such as clothing.
HUGS was done in collaboration with @mkocab_, Rick Chang, James Gabriel, and @OncelTuzel. （from @anuragranj）