CUDA Voxelization (PyTorch + VisPy)

Minimal playground for GPU voxelization of KITTI point clouds and quick 3D visualization.

Prereqs

• Python 3.10+
• CUDA-capable GPU + CUDA drivers
• uv (or pip) to install deps

Install

uv sync

Install a matching CUDA PyTorch before building the extension (torch is not in build-system deps):

pip install --extra-index-url https://download.pytorch.org/whl/cu124 torch torchvision torchaudio

Build the CUDA extension in-place (works with uv or plain python):

uv run setup.py build_ext --inplace
# or
python setup.py build_ext --inplace

Visualize a folder of KITTI .bin files

uv run vis_simple.py /path/to/velodyne

Keys: Right/N/Space/Enter → next file, Left/B → previous. Close the window to quit.

What it does

• Loads KITTI-format point clouds (x, y, z, intensity)
• Runs CUDA voxelization (voxelization/ops.py binding)
• Colors points/voxels by distance from the origin using a jet colormap
• Renders with VisPy turntable camera

Note

The wheel/extension is built via the included CUDA source in voxelization/csrc. Use the provided uv.lock/pyproject.toml for reproducible builds.