Synthetic Dataset Generator

Authors:

• Stanislava Fedorova
• Alberto Tono
• Meher Shashwat Nigam
• Jiayao Zhang
• Amirhossein Ahmadnia
• Cecilia bolognesi
• Dominik L. Michels

This is a tool that generates a dataset of synthetic buildings of different typologies.

Arxiv Website Samples

The generated data includes:

• Mesh files of generated buildings, .obj format
• Rendered images of the mesh, .png format
• Rendered segmentation masks, .png format
• Depth annotation, .png and .exr format
• Surface normals annotation, .png format
• Point cloud files, .ply format (the number of points by default is 2048, can be changed in dataset_config.py)

How To Use

• Install Blender>=2.90. After installation make sure to add blender as an Environment variable.
• Download the package as a .zip file or:

git clone https://github.com/CDInstitute/CompoNET

*Navigate to the Building-Dataset-Generator folder.

pip install -r requirements.txt

To create completely synthetic buildings use:

run.bat

Or:

blender setup.blend --python dataset.py

Unfortunately, it is not possible to use Blender in background mode as it will not render the image masks correctly.

Note: all the parameters related to the dataset (including any specific parameters for your buildings (e.g. max and min height / width / length)) are to be provided in dataset_config.py. Default values adhere to international standards (min) and most common European values (max):

• minimum height 3m
• minimum length and width 6m
• maximum length, width, height 30 m Other values to set:
• number of dataset samples
• building types
• component materials
• rendered image dimensions
• number of points in the point clouds
• paths to store the generated data
• option to save the .exr files

Annotation structure

{'img': 'images/0.png', 'category': 'building', 'img_size': (256, 256), '2d_keypoints': [], 'mask': 'masks/0.png', 'img_source': 'synthetic', 'model': 'models/0.obj', 'point_cloud': 'PointCloud/0.ply', 'model_source': 'synthetic', 'trans_mat': 0, 'focal_length': 35.0, 'cam_position': (0.0, 0.0, 0.0), 'inplane_rotation': 0, 'truncated': False, 'occluded': False, 'slightly_occluded': False, 'bbox': [0.0, 0.0, 0.0, 0.0], 'material': ['concrete', 'brick']}

Performance

We ran the dataset generation algorithm for 100 model samples with different input parameters on Windows 10 OS on CPU and GPU using AMD Ryzen 7 3800-X 8-Core Processor and GeForce GTX 1080. Here we report the results for the multiview generation (3 views per model):

Citation

Bibtex format

@inproceedings{fedorova2021synthetic,
      title={Synthetic 3D Data Generation Pipeline for Geometric Deep Learning in Architecture}, 
      author={Stanislava Fedorova and Alberto Tono and Meher Shashwat Nigam and Jiayao Zhang and Amirhossein Ahmadnia and Cecilia Bolognesi and Dominik L. Michels},
      year={2021},
}

Generated Image Samples