BuildNet3D - Synthetic 3D Building Dataset

This dataset contains diverse synthetic 3D building models, each accompanied by 200 multi-modal images and corresponding ground-truth building characteristics. The building models are classified into three categories: synthetic skyscrapers, synthetic L-shaped buildings, and realistic, geometrically complex structures.

Dataset Structure

• Image Dataset (Each type includes 4 building models)
  • Color Images (images/)
  • Depth Maps (depths/)
    • Multiplied by a scaling factor of 1000
    • Higher precision provided for realistic building models
  • Normal Map (normals/ - Realistic Building only)
  • Instance Masks (instances/ - Realistic Building only)
    • Colored masks corresponding to instances and individual facades
  • Semantic Masks
    • semantic_masks/ - Pixel-wise object IDs representing object classes
    • semantics/ - Colored masks corresponding to semantic categories
  • Segmentation Information (segmentation_data.json)
    • Mapping from category IDs to their corresponding semantic colors
  • Metadata (meta_data.json)
    • Camera pose and intrinsic parameters for each frame

For more information about our characteristic estimation framework, please refer to our publication at link.  

The source codes to reproduce the results are shared at link.

Please cite the following article:

[1] Xu, C., Mielle, M., Laborde, A., Waseem, A., Forest, F., & Fink, O. (2025). Exploiting semantic scene reconstruction for estimating building envelope characteristics. Building and Environment, 112731. doi: https://doi.org/10.1016/j.buildenv.2025.112731