Source code of building iShape synthetic data by Blender

Visualization of generated dataset

About iShape dataset: https://ishape.github.io/

Source code of building iShape-Branch, Fence, Log, Hanger, and Wire.

We provide two methods to run this code:

1. Docker: We are highly recommend this solution. All you need is installing docker and run a single command line.
2. Install from scratch: A little bit complex. If you get any problem when install from scratch, you can also refer to Dockerfile which includes details installation instructions.

▮ Docker

Build dataset by one command line:

mkdir synthetic_ishape_dataset

# Docker size of diyer22/ishape is about 15GB
docker run -v `pwd`/synthetic_ishape_dataset:/synthetic_ishape_dataset diyer22/ishape

▮ Install from scratch

Require:

• Ubuntu (tested on 18.04)
• Python >= 3.7
• Blender >= 2.90

Steps:

1. Download and install Blender
2. Install bpycv for Blender's bundled Python
3. Prepare code and asset:

mkdir ishape_dataset && cd ishape_dataset
# prepare code and source asset
git clone git@github.com:iShape/build_synthetic_ishape.git
git clone git@github.com:iShape/source_asset.git

4. Prepare background:

   • python build_synthetic_ishape/tool/download_background_hdri.py
   • Which will download hdr file from HDRI Haven to build_synthetic_ishape/source_asset/shared/hdri(about 11GB)

5. Synthesis dataset by Blender:

   • cd build_synthetic_ishape

   • You can build whole iShape synthetic dataset by:

     • python build_synthetic_ishape.py

   • Or build some sub-datsets like:

     • blender --background --python branch.py -- DIR ../synthetic_ishape_dataset/branch/train/ IMG_NUM 2000

▮ Others

Dataset file struct: tree synthetic_ishape_dataset/branch/train/ -L 1

synthetic_ishape_dataset/branch/train/
├── depth
├── image
├── instance_map
├── vis # Visualization
└── ycb_6d_pose

Visualization: Open synthetic_ishape_dataset/branch/train/vis/100.jpg:

Generate COCO style dataset (optional):

python build_synthetic_ishape/tool/instance_map_to_coco.py \
  --dirr synthetic_ishape_dataset/branch/train --mask_encoding rle # Or --mask_encoding poly

ls synthetic_ishape_dataset/branch/train/coco_format/
# instances_train2017.json
# train2017