Challenges

Submission Tools

3D Reconstruction - CHI3D

The purpose of this challenge is to advance the state of the art in 3D human reconstruction of close contact interactions. We standardize the evaluation protocol and metrics to facilitate the comparison of different research methods.

Our evaluation server accepts multiple formats of the reconstructions: The following evaluation metrics are computed:
  • MPJPE: Mean Per Joint Position Error
  • MPJPE-PA: Mean Per Joint Position Error after Procrustes Analysis
  • Translation Error computed on the pelvis joint
  • GHUM MPVPE: Mean Per Vertex Position Error in the GHUM mesh format
  • GHUM MPVPE-PA: Mean Per Vertex Position Error after Procrustes Analysis in the GHUM mesh format
  • SMPLX MPVPE: Mean Per Vertex Position Error in the SMPLX mesh format
  • SMPLX MPVPE-PA: Mean Per Vertex Position Error after Procrustes Analysis in the SMPLX mesh format
Evaluation is performed against both persons. Person matching (between predictions and ground truth) is performed per-frame, using 2d bounding boxes, so the order of predicted persons in each sequence does not matter.

The TEST set consists of only one random camera viewpoint per interaction scenario. Each prediction has to be provided in the coordinate system of the provided camera.

We require predictions only on a subset of frames, equally distanced and provided in the Template File. All evaluation metrics are averaged and reported over (1) all frames that were sampled, as well as only over the (2) contact frames (one interaction-contact frame per sequence, as reported in the paper).

Please take into account that there is a fixed time window of 3 days between consequent submissions! To avoid wasting submissions, please validate their format using the Format Validation Code and feel free to render the predictions with the provided Visualization Code. 		Template File
Format Validation Code
Visualization Code