• an implementation of a classical algorithm for segmenting surfaces in point clouds, called Region Growing Segmentation (RGS). It is based on RGS
• a deep learning pipeline for segmenting faces in 3D point clouds

Create a virtual environment "venv_deepseg" with all required dependencies as follows:

• Clone this repository
• Use a shell to execute following commands:
  • cd deepsegmentation
  • conda create -y -n venv_deepseg python=3.6.8 anaconda
  • conda activate venv_deepseg
  • sh requirements.txt

• Root folder:
  • Contains all available jupyter notebooks
  • Contains shared functions
  • Contains configuration files:
    • config.json: Base file for network configuration
    • generator_config.py: Dictionary to adjust the generated point clouds
• Classic_RGS: Contains all program code that is only linked with the RGS algorithm. Also input and output files related to RGS are stored here. Note: The point cloud generator is not configured to store its output in this folder. Point clouds need to be manually shifted to the input folder.
• data: Contains program code for the point cloud and patch datasets. Contains all data that is linked with the deep learning pipeline.
  • input_data: When a dataset is generated, data is automatically stored here
  • network_data: Configuration files and network configurations are stored here during training
  • output_data: Plots during training, testing results and prediction results are stored here.
  • temp_data: Temporary data is stored here and continuously overwritten. Used e.g. when running the voxelization notebook.
  • tensor_board_logs: Not used.
• Network: Contains program files linked to the deep learning pipeline: Preprocessing, Network, Postprocessing
• PolyhedronGenerator: Contains the program code for the polyhedron generator.
• utils: Contains utility program code

• Region Growing Segmentation:
  • Open the Jupyter notebook "Classic_RegionGrowingSegmentation".
  • Make sure that the input path is correctly set.
  • After running the notebook, find the files in the specified output path.
• Deep Learning Pipeline:
  • You can evaluate existing models via the notebooks "Predict" or "Evaluate". The difference is, that Evaluate has access to the ground truth data and returns more information.
  • You can train a new or existing model using the notebook "Training".
    • By setting an experiment name, it is assured that results are separated from other configurations.
    • To overwrite an existing experiment, uncomment the 'delete' line.
    • To continue training an existing experiment, uncomment the 'create_retrain_experiment' line. Make sure that the 'delete line' is commented.
  • Datasets will automatically be created in these notebooks, if not available. Make sure that "generator_config.py" is set as intended.
  • All results are stored in data/output_data
  • After training is finished, you can call the notebook "Read_Training_Results" which will display results on the validation set and automatically generate an html version for archiving.
• Independent notebooks:
  • Notebook "PolyhedronGenerator": You can use this to try the polyhedron generator. Data will be stored in data/temp_data
  • Notebook "Voxelization": You can observe how the patches in a point cloud look and how a patch looks compared to its voxelized version. Data will be stored in data/temp_data

The following requirements are all satisfied via the installation procedure given above.

• Python 3.6.8
• Pytorch 1.1.0
• numpy 1.15.4
• numba 0.43.1
• plotly 3.10.0
• matplotlib 3.0.2
• scikit-learn 0.21.2
• scipy 1.3.0
• tensorboard 1.14.0
• tensorboardx 1.7
• jupyter 1.0.0
• python-constraint 1.4.0
• cvxpy 1.0.24