Learning Complex 3D Human Self-Contact

Monocular estimation of three dimensional human self-contact is fundamental for detailed scene analysis including body language understanding and behaviour modeling. Existing 3d reconstruction methods do not focus on regions in consequently recover configurations that are either far from each other or self-intersecting, when they should just touch. This leads to perceptually incorrect estimates limits impact those very fine-grained domains where models expected play an important role. To address such challenges we detect design losses explicitly enforce it. Specifically, develop a model Self-Contact Prediction (SCP), the surface signature self-contact, leveraging localization image, during both training inference. We collect two large datasets support learning evaluation: (1) HumanSC3D, accurate motion capture repository containing 1,032 sequences with 5,058 contact events 1,246,487 ground truth poses synchronized images collected multiple views, (2) FlickrSC3D, 3,969 images, 25,297 surface-to-surface correspondences annotated image spatial support. also illustrate how more expressive reconstructions can be recovered under constraints present monocular detection face-touch as one applications made possible by models.