Computer Vision Examples Monorepo

A Pixi workspace of computer vision applications, reusable libraries, and source-built dependencies. Runnable projects live under packages/ and compose shared features with package-specific dependencies in the root pixi.toml. Most runnable packages have a production environment (<name>) and a development environment (<name>-dev) with the same dependency solve.

The workspace platforms are linux-64, linux-aarch64, and osx-arm64, plus the named linux-64-cuda13 and linux-aarch64-cuda13 platforms that carry the CUDA 13.0 virtual packages for GPU environments. Each feature opts into its platforms explicitly, so most GPU packages are Linux-only; check the package README and its feature in pixi.toml before installing on macOS. See the "Platforms & lockfile" section of AGENTS.md for the full rules.

Quick start

1. Install Pixi, then restart your shell:

   curl -fsSL https://pixi.sh/install.sh | sh
   

2. Clone the workspace and install one development environment:

   git clone https://github.com/rerun-io/examples-monorepo.git
   cd examples-monorepo
   pixi install -e monoprior-dev --frozen
   

3. Run that package's demo task from the repository root:

   pixi run -e monoprior-dev --frozen monoprior-relative-depth
   

Choose another package from the map below, then use its README or pixi task list -e <name> to find its demo task. Tasks with download dependencies fetch their example assets on first run.

Optional: direnv

The root .envrc activates dev. Package directories that contain their own .envrc default to <name>-dev, so after installing direnv you can opt in per directory:

cd packages/monoprior
direnv allow
pytest -q

Set PIXI_ENV in a gitignored .envrc.local to override a directory's default.

Package map

Build-only and vendored packages do not necessarily have standalone root environments or demo tasks; their role is described in the second column.

Conventions

• Root-managed runnable packages generally use <name> for production and <name>-dev for development. The [environments] table in pixi.toml is the source of truth.

• Development environments expose the canonical lint, typecheck, deadcode, and tests tasks:

  pixi run -e robocap-slam-dev --frozen lint
  pixi run -e robocap-slam-dev --frozen typecheck
  pixi run -e robocap-slam-dev --frozen deadcode
  pixi run -e robocap-slam-dev --frozen tests
  

• Prefer pixi run --frozen and pixi install --frozen while dependencies are unchanged. Omit --frozen only when intentionally updating the solve and lockfile.

• List the tasks available in an environment with pixi task list -e <name> or pixi task list -e <name>-dev.

• In-repo packages depend on each other by name in [project].dependencies and pin that name to the sibling's path with [tool.uv.sources] in their own pyproject.toml; the root manifest installs the same paths editable. The pin keeps a PyPI namesake (there is one for monopriors) out of the solve. pixi-build source packages, which asmk, dpretrieval, and mast3r already are, are the eventual replacement once that preview stabilizes.

For agent and contributor documentation — architecture, platform rules, adding a new package, Rerun version policy, and known issues — see AGENTS.md.