Subjects in self-contact

  • TRAIN set: 4 subjects
  • TEST set: 2 subjects

In each recording, the subject is motion tracked with a marker-based motion capture system (Vicon).

Multiple cameras

  • 4 different views
  • 900x900 resolution
  • 50 fps
  • Camera parameters:
    • extrinsics
    • intrinsics for 2 different camera models (one assuming image distortion, one ignoring it)
  • The TEST set consists of only one random frontal camera viewpoint per sequence, to avoid simplifying the 3D Reconstruction challenge through multi-view triangulation/optimization.

Various self-contact actions in various scenarios

Each subject performs the following actions.

  • standing (116 scenarios)
  • sitting on the floor (20 scenarios)
  • interacting with a chair (36 scenarios)

GHUM and SMPLX meshes

  • Ground-truth, well-alligned mesh - obtained by fitting the GHUM model to accurate 3d markers, multi-view image evidence and body scans
  • We retarget the GHUM meshes to the SMPLX topology and provide pose and shape parameters for both
  • 50 fps

3D skeletons

  • Ground-truth 3d skeletons with 25 joints (including the 17 Human3.6m joints)
  • 50 fps

Self-contact annotations

Each of the 1032 recordings contain:

  • 1 video timestamp where self-contact is established
  • the self-contact signature annotation (multiple correspondences on the body surface between two triangle ids / vertex ids / region ids)
  • the self-contact image support annotation (the projection of each self-contact vertex correspondence in the image)

Due to the 4 viewpoints, this amounts to 4128 triplets of images, self-contact signatures and self-contact image support.