Skip to content

Latest commit

 

History

History
138 lines (96 loc) · 3.21 KB

challenge_1_track_2.md

File metadata and controls

138 lines (96 loc) · 3.21 KB

Challenge 1 - Track 2: Recovery of Fine Details in Human Scans

Given a partial human body scan, X, the goal is to recover the complete scan, Y.

This challenge uses the 3DBodyTex 2 dataset.

Training data

Provided:

  1. A set of simulated complete scans, Ys.
  2. Additional metadata: The 3D positions of detected body landmarks.

Not provided:

  1. The partial scans, X. They are generated by the participants.

The body landmarks are used to generate the partial data. They may be freely exploited during training but they are not provided for the final evaluation.

Both meshes are aligned and in the same frame of reference.

Evaluation data

Provided:

  1. A set of partial scans, X. These are generated by the organisers from Ys.

Not provided:

  1. The ground-truth scans, Y = Yf. (They will be released after the competition.)
  2. Additional metadata.

Evaluation

Considered for the evaluation:

  • the shape,
  • ears, hands and feet.

Not considered for the evaluation:

  • the texture,
  • the rest of the body.

See evaluation for the definition of the metric.

The evaluation is performed on the ears, the hands and the feet (areas with fine details). The rest of the body is ignored. The hand region starts at the wrist. The foot region starts at the ankle.

The texture is ignored as it is less reliable on the areas with fine details on the synthetic data.

Submission format

The predicted complete mesh, Y', should be in .obj or .npz (see formats).

The mesh colour information should be either:

  1. a single texture atlas,
  2. or RGB colour stored as vertex attributes.

A mixture of vertex colour attribute and texture mapping is not allowed.

The mesh geometry may be different than the input, X.

The submission must be a .zip archive that unfolds in this directory structure:

  eval/
    <scan_name>/
      <scan_name>-completed.(npz|obj)
    .../

where <scan_name> correponds to the name of the input scan.

Directory structure of the dataset

The dataset has the following directory structure:

  train/
    <scan_name>/
      fitted_textured.npz
      fusion_textured.npz
      landmarks3d.txt
    .../
  test/
    <scan_name>/
      fitted_textured.npz
      fusion_textured.npz
      landmarks3d.txt
    .../
  eval/
    <scan_name>/
      fitted_textured.npz
      fusion_textured.npz
      <scan_name>-synthetic-partial.npz
      landmarks3d.txt
    .../

For each scan, there is one subdirectory with a unique <scan_name>, e.g. 170926-001-fitness-run-sbt1-d8f6-low-res-result. The files are:

  • fitted_textured.npz = Yf: The reference mesh with details. It is obtained by fitting the SMPL-X body model to a body scan and transferring the texture.
  • fusion_textured.npz: Ys. The simulated body scan. This is a textured mesh obtained by simulating the scanning process in software on Yf. It is less detailed and contains artefacts similar to a real 3D scan.
  • <scan_name>-synthetic-partial.npz = X: Partial scan generated from Ys.
  • landmarks3d.txt: The 3D positions of the detected 3D landmarks. They are common to both meshes.

See formats for the mesh and landmark data formats.