PointCloud Workflows is a GPU-first point-cloud segmentation toolkit for NVIDIA RTX 4070 systems, with a portable CPU fallback. It combines synthetic data generation, dataset validation, voxel preprocessing, training, evaluation, and prediction. The project originated in LiDAR segmentation and BIM research and provides reusable public dataset contracts and reproducible workflows.
This is the documentation source for the public PointCloud Workflows successor.
It derives from RPTU AI Toolkit and starts with clean project history. Install the
pointcloud-workflows distribution from this checkout and use pcw or
python -m pointcloud_workflows. The old commands/imports remain aliases;
saved artifact formats are unchanged. See the migration guide
and project lineage.
The public RPTU guide documents the shared baseline. No PointCloud Workflows documentation site has been published; see publication status.
WarpConvNet's MinkUNet18 provides spatial segmentation on NVIDIA GPUs. The PyTorch voxel MLP provides a lightweight CPU fallback and can also run on CUDA. Both use the same dataset contracts, CLI, original point mapping, and model-bundle format. Saved models retain their architecture; a WarpConvNet model requires CUDA, while a PyTorch model can move between CPU and CUDA.
Start with the GPU and CPU quickstart, then read the CLI reference, dataset contract, and model guide. The training guide explains model selection; test and evaluation covers metrics and point-label exports.
The toolkit includes small seeded synthetic scenes so the complete workflow can run without an external dataset account. These scenes demonstrate behavior, not real-world segmentation quality. Train a fresh model on your own labeled data for an application; original research datasets and weights are not distributed.
Training supports epoch checkpoints and --resume, with durable attempt records
and verified completed retries. The roadmap continues with public-data
recipes, spatial chunking, releases, Helm Jobs, and Fleet examples. Use pcw for the current workflow; research scripts remain in the original RPTU repository.
The NumPy-only PointCloud Core integration supplies the shared array operations while PCW retains dataset formats and workflow execution. Its standalone example can run independently of PCW.
For integration from another program, use the versioned JSON execution protocol. It includes structured failures and read-only run inspection while preserving ordinary CLI and saved-artifact behavior.
Public source: ajavanma/PointCloud-Workflows on GitHub. The project is MIT licensed.