3D Scan of Subjects with Hand and Face Articulation
The EHF dataset (Expressive Hands and Faces) contains 100 curated frames of one subject in minimal clothing, performing a variety of body poses, including natural finger articulation, as well as some facial articulation and expressions.Each frame includes the following time-synchronized modalities: a full-body RGB image, a JSON file with 2D features detected with OpenPose (body joints, hand joints, facial features), a 3D scan of the subject, a 3D SMPL-X alignment (3D mesh) to the above scan, functioning as pseudo ground-truth.The pseudo ground-truth meshes facilate using a vertex-to-vertex (v2v) error metric. This is a stricter metric than the common paradigm of 3D joint error that does not capture surface errors and rotations along the bones.The SMPL-X model and SMPLify-X code to reconstruct 3D humans from a single RGB image are available.