toys4k

Toys4K 3D Object Dataset

This directory contains instructions on how to download the 3D assets for the Toys4K dataset and render them into training data.

Dependencies

Blender

Download Blender 2.80 for your operating system from the official Blender release.

For MacOS the default installation path is /Applications/Blender/blender.app/Contents/MacOS/blender.

For Linux it involves extracting a blender-2.80-linux-glibc217-x86_64.tar.bz2 which makes the path <path to extracted tarball>/blender-2.80-linux-glibc217-x86_64/blender.

Depending on your OS, add the appropriate path in the rendering Bash scripts renderer/render_demo.sh and renderer/render_toys.sh under the variable blender_path.

Python

The simplest way to make sure all the right Python dependencies are in place is with conda.

conda env create -f environment.yml
conda activate blender_datagen

View Demo Renders - Render Your Own Data

To generate a set of demo renders to test your install and our code run the following commands

cd renderer
bash render_demo.sh

After this has completed, you should see the following images under renderer/demo_output/airplane/airplane_010/image_output

Downloading Toys4K

To download the dataset, please fill out the following form. You will get access to:

toys4k_blend_files.zip - blend files to use for rendering
toys4k_obj_files.zip - extracted obj files without materials to use for 3D geometry processing-type tasks
toys4k_point_clouds.zip - extracted surface point clouds and normals
toys_blend_sample_renders.zip - sample renders from the blend files
toys_obj_sample_renders.zip - sample renders from the obj files

Creating Training Data

Toys 4K

1. Download the Toys .blend files archive as specified in the instructions above and unzip it.
2. Set the appropriate paths in renderer/render_toys.sh and run bash render_toys.sh

ModelNet40

1. Download the manually aligned modelnet and unit

wget https://lmb.informatik.uni-freiburg.de/resources/datasets/ORION/modelnet40_manually_aligned.tar

2. Convert to .obj files (required for Blender rendering). Expect this to take a few hours to complete.

python util_scripts/convert_to_obj.py --src_path=<path to untarred directory> --dest_path=<path to output the converted files>

3. Set the appropriate paths in renderer/render_modelnet.sh and run bash render_modelnet.sh

ShapeNetCore.v2

1. Download ShapeNetCore.v2 from the official webpage.
2. Set the appropriate paths in renderer/render_shapenet.sh and run bash render_shapenet.sh

Notes

• Some objects have been removed from the inital set of renders/point clouds to allow for this release.
• You may need to explicitly enable CUDA and GPUs at Edit/Preferences/System in Blender in order to do GPU-based rendering.