BuildNet3D

Official code for "Exploiting Semantic Scene Reconstruction for Estimating Building Envelope Characteristics" (Building and Environment 2025)

3D Semantic Building Reconstruction

This repository extends nerfstudio (v1.1.3) to reconstruct building envelope surface meshes with appearance, geometry, and semantic properties using SDF-based representations.

It also provides baseline models, such as nerfacto (NeRF-based), neus(SDF-based), and splatfacto(Gaussian-Splatting-based), for pure RGB reconstruction.

Installation

# Clone repository
​git clone --recursive git@github.com:EPFL-IMOS/buildnet3d.git

# Install dependencies
export TORCH_CUDA_ARCH_LIST="<YOUR_ARCH>"
pip install --no-binary :all: gsplat==1.0.0 # if you need gaussian splatting
cd buildnet3d/nerfstudio
pip install -e .

# Install tinycudann
TCNN_CUDA_ARCHITECTURES=<YOUR_ARCH> \
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch 

Tested on A100 GPU with Python 3.10, PyTorch 2.0.1, and CUDA 11.8. For splatfacto, use PyTorch 2.1.2 and torchvision 0.16.2 with CUDA 11.8.

Usage

# Reconstruction
python buildnet3d/scripts/reconstruct.py --model-type <MODEL_TYPE> --data <DIR_TO_DATA>

# Evaluation
python buildnet3d/scripts/evaluate.py --load-config <MODEL_CONFIG>

The train/test split is defined in buildnet3d/data/buildnet_dataparser.py.

Multi-Modal Imageset Generation

This repository uses BlenderProc to generate multi-modal image data from 3D building models. The rendered outputs include RGB, depth maps, surface normals, semantic labels, and instance segmentations. We implement a simple rule-based sampling method to randomly place camera viewpoints while ensuring the entire object remains within the view. More details are provided here.

The generated buildnet3d image dataset is available here.

Citation

If you find this repository or the associated dataset useful, please cite:

@article{XU2025112731,
      title = {Exploiting semantic scene reconstruction for estimating building envelope characteristics},
      journal = {Building and Environment},
      volume = {275},
      pages = {112731},
      year = {2025},
      issn = {0360-1323},
      doi = {https://doi.org/10.1016/j.buildenv.2025.112731},
      url = {https://www.sciencedirect.com/science/article/pii/S0360132325002136},
      author = {Chenghao Xu and Malcolm Mielle and Antoine Laborde and Ali Waseem and Florent Forest and Olga Fink},
}