How vlar-group/growsp works
Project type and purpose
• This is an ML research codebase (“GrowSP”) for unsupervised semantic segmentation of 3D point clouds.
• It provides end-to-end pipelines for data preparation, initial superpoint graph creation, model training, and evaluation.
Relationships and interactions
• The data_preparation scripts write to ./data/S3DIS/{input,initial_superpoints}
• The dataset loader reads from these folders
• The training script writes model artifacts to ./ckpt/S3DIS
• The evaluation script reads from ./ckpt/S3DIS and ./data/S3DIS/{input,initial_superpoints}
Architectural patterns & design principles
• Modular separation of concerns: data prep, dataset abstraction, model definition, training loop, evaluation
• Use of a shared utils library (lib/) to avoid duplication
• Explicit folder structures for each dataset to enable multi-dataset support
Technologies & frameworks
• Python 3, Conda environment (env.yml)
• PyTorch + MinkowskiEngine for sparse 3D convolutions
• Standard open-source libraries (numpy, scipy, yaml, etc.)