vision3d 0.3.0

A 3D extension of torchvision.

vision3d

This library is a 3D extension of torchvision, providing datasets, tensor types, transforms, ops, metrics, and visualization utilities for 3D perception tasks.

Documentation is available at vision3d.dev.

[!WARNING] vision3d is in active early development. The API may change without notice and documentation may be incomplete.

Requirements

• Python 3.12 or newer.
• PyTorch 2.10 or newer.
• Recommended: A CUDA-capable NVIDIA GPU for GPU execution.
• For building from source: the CUDA toolkit matching your PyTorch build.

Installation

vision3d is published on PyPI as a pre-built wheel and sdist. The wheel is built against the LibTorch Stable ABI and statically links the CUDA runtime, so one wheel works for any Python 3.12+, torch 2.10+, and any NVIDIA driver that supports CUDA 12.8 or newer (Linux driver ≥ 570).

From PyPI

We recommend using uv as your package manager:

uv add vision3d

Or with pip:

pip install vision3d

From source

Clone the repository and sync the environment:

git clone https://github.com/yeetypete/vision3d.git
cd vision3d
uv sync --all-extras

uv sync compiles the C++/CUDA extension as part of installing the project. On machines where CUDA is installed but no GPU is visible (for example, inside containers), force a CUDA build with:

FORCE_CUDA=1 TORCH_CUDA_ARCH_LIST="12.0+PTX" uv sync --all-extras

[!NOTE] TORCH_CUDA_ARCH_LIST selects which NVIDIA compute capabilities to compile CUDA kernels for (e.g. 12.0 for RTX 50-series). See the PyTorch docs for the full syntax.

To produce a wheel locally:

uv build

By default uv build resolves torch from PyPI, which currently ships the cu130 variant. If your local CUDA toolkit is a different major version, point uv at the matching PyTorch wheel index instead:

uv build --index https://download.pytorch.org/whl/cu128

Replace cu128 with whatever CUDA major version your installed CUDA toolkit ships, e.g. cu130, cu132.

Extras

• nuscenes: pulls in nuscenes-devkit for the nuScenes dataset loader.
• viz: pulls in rerun-sdk for the visualization utilities in vision3d.viz.

Request them at install time, for example: uv add 'vision3d[nuscenes,viz]'.

Contributing

Contributions are welcome! See CONTRIBUTING.md for how to get started.

License

vision3d is released under the BSD 3-Clause License.