Skip to content

Latest commit

 

History

History
172 lines (127 loc) · 5.25 KB

dataset_3dbodytex2.md

File metadata and controls

172 lines (127 loc) · 5.25 KB
Raw

3DBodyTex 2

3DBodyTex 2 contains thousands of 3D textured human scans. It is an extension of 3DBodyTex.

Sample scans from the 3DBodyTex 2 dataset Subset of 3DBodyTex 2.

Content

  • 3D human scans (Y):
    • acquired with the Shapify Booth,
    • many subjects,
    • varied poses,
    • varied casual clothing,
    • standard tight-fitting clothing,
    • 3D watertight meshes with a high-resolution texture atlas;
  • 3D landmarks:
    • body joints and landmark points on the surface,
    • estimated automatically;
  • Fitted textured SMPL-X model (Yf):
    • SMPL-X is fitted to all the raw scans Y with close-fitting clothing to obtain Yf,
    • pose and shape parameters,
    • fixed template mesh (SMPL-X),
    • texture mapped from the raw scans;
  • Simulated 3D scan acquisitions (Ys):
    • the process of the Shapify Booth is simulated in software,
    • the acquisition is simulated on the fitted SMPL-X models Yf to obtain Ys;
  • Synthetic pairs (Yf, Ys) of (detailed ground truth, acquisition):
    • Yf is a synthetic groud truth with detailed shape and texture,
    • Ys is a synthetic scan acquisition with typical artefacts,
    • for scans with close-fitting clothing only.

Usage in the SHARP Challenge 1

Track 1: Recovery of large regions

  • The ground truth is Y.
  • The partial scans are generated from Y.
  • The 3D landmarks are provided at training time only, not at evaluation time.

Track 2: Recovery of fine details

  • The ground truth is Yf.
  • The partial scans are generated from Ys.
  • The 3D landmarks are provided at training time only, not at evaluation time.

Detailed description and statistics

Clothing

casual fitness
casual fitness

Poses

A U run scape free
... unconstrained

Synthetic detailed ground-truth and simulated acquisitions

The goal is to obtain a ground-truth shape preserving the details of smaller-scale regions (e.g. fingers and feet). This cannot be acquired systematically with a body scanner. Instead, the template mesh of a body model (SMPL-X) is used as a reference for the fine-scale shape details.

The procedure to synthesise the data is:

  1. Y = input: raw scan in close-fitting clothing,
  2. Yf = fit Y: fit the body model to Y and transfer the texture to obtain Yf,
  3. Ys = acquire Yf: simulate the acquisition process on Yf to obtain Ys.

The pair (Yf, Ys) is a synthetic pair of (acquired, ground-truth) shape that can be used for supervision.

scan (Y) fitted (Yf) simulated acquisition (Ys)
fitness fitted simulated

Encoding formats and conventions

File formats

Meshes are in the .npz format. Landmark positions are in a tabular text format. See formats.

Orientation

The up direction is the y axis. The ground is the x-z plane. Most scans are facing the -z direction, but not all.

Units

Vertex positions are in metres.

3D body landmarks

There are 67 body landmarks detected automatically on each scan. They are provided to generate the partial data but may also be used for training. They comprise standard body joints and other keypoints on the body (eyes, nose, ears...). The detection of most landmarks is stable except for the finger joints and finger tips.

Below, <side> is either left or right.

Body joints (20)

landmark name
ankle_<side>
elbow_<side>
heel_<side>
hip_<side>
hip_middle
knee_<side>
neck
shoulder_<side>
toe_1_<side>
toe_5_<side>
wrist_<side>

Face landmarks (5)

landmark name
ear_<side>
eye_<side>
nose

Finger joints and tips (42)

<knuckle> is one of base, middle, top, tip.

landmark name
finger_baby_<knuckle>_<side>
finger_index_<knuckle>_<side>
finger_middle_<knuckle>_<side>
finger_ring_<knuckle>_<side>
finger_thumb_<knuckle>_<side>
hand_base_<side>