Omniglot3D

Overview

This is 3D meshes of letters created out of Omniglot dataset.
Each letter represented as .glb file. Meshes are estimated out of svg images.

Use

python 3.9

Enviroment setup:

# CONDA #
conda create --name omniglot3d python=3.9
conda activate omniglot3d
pip install -r requirements.txt

# VENV #
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt install python3.9
sudo apt update
sudo apt install python3.9 python3.9-venv
python3.9 -m venv omniglot3d_venv
source omniglot3d_venv/bin/activate
pip install -r requirements.txt

From 2D to 3D pipe:

0. Run get_data.sh to load data from original repo.
   Or put original "Omniglot" into ./data/ folder
   you need extract images_background.zip.
1. image2svg.py to generate svg conturs for extrusion meshes.
2. Run python svg2mesh.py to generate “broken” glb files (something is wrong with trimesh glb export for normals) they are fine, but they need to be resaved in blender.
3. Run blender --background --python fix_glbs.py to resave glb file. (Blender v.3.0.1) That might take some time.

Final .GLB meshes will be in ./processed_data/fixed_glbs with same folder structure as original Omniglot.

If you just want data there is two archives in ./processed_data for train and evaluation respectfully. Use it as you see fit.

Usecase

Syntetic 3D data of generic(but human) origin for detection and localization tasks in 3D.

Known issues

• Seems like meashes rotated 180 degees on y-axis. (Upsidedown)
• Topology is not great, maybe marching cubes pass is in order. (in blender pass, but it will take even longer)

Thanks

I wanna acknowledge again original Omniglot repo and all people involved as well as an original website.

• https://www.omniglot.com
• https://github.com/brendenlake/omniglot