A workshop and challenge in conjunction with ECCV 2020.
The goal of the competition is to recover a reference textured scan Y from a
partial version of it X.
This is a completion task of 3D shape and texture.
The data consists in pairs (X, Y) generated from two
datasets of 3D textured scans:
- 3DBodyTex 2: a dataset of 3D human scans (extending 3DBodyTex).
- 3DObjectTex: a dataset of 3D scans of generic objects.
There are two challenges, one per dataset.
The ground-truth shapes, Y, are the raw scans (or the fitted body models in
a variant of the challenge on human scans).
The partial scans, X, are generated synthetically from the ground-truth
shapes, Y.
The idea is to simulate the partial acquisition produced by a 3D scanner, e.g.
a hand-held scanner.
For training, only the ground-truth shapes, Y, are provided.
The partial scans, X, are generated by the participants.
Some example routines are provided.
Custom routines may be used if reported.
For evaluation, a set of partial shapes, X, is provided to the
participants.
The corresponding ground-truth shapes, Y, are kept secret.
The participants estimate a complete shape, Y', from X.
The submissions are evaluated quantitatively by computing the
surface-to-surface distance between Y and Y'.