BEHAVE: the first dataset with accurate 3D human-object reconstructions paired with images. BEHAVE also includes a reliable neural reconstruction-fitting method to automatically obtain SMPL + object fits from multi-view depth. #CVPR2022.
The field of human analysis is shifting towards human-object and scene interaction, but obtaining reliable and diverse data is hard. If you are working on 3D human pose and shape from images / video / depth/ multi-view, interaction modelling, human-scene interaction, motion forecasting, check out BEHAVE!
Arxiv: https://arxiv.org/abs/2204.06950
Dataset: http://virtualhumans.mpi-inf.mpg.de/behave
#cvpr2022 #3D HPS #3DHumans #3Dinteraction
