Docker#

Docker provides a convenient way to build, install and run applications isolated from the rest of your system. You do not need to change software versions on your system or install new software, except the Docker engine itself.

First, install Docker following the official guide. Also, complete the post-installation steps for Linux. Make sure that docker can be executed without root privileges. To verify Docker is installed correctly, run:

# You should be able to run this without sudo.
docker run hello-world

Install and run Open3D apps in docker#

You can install and run Open3D applications from a docker container.

For Python application, you will need to install a minimum set of dependencies. For more details please see this issue. A minimal Dockerfile for Python applications looks like this:

# This could also be another Ubuntu or Debian based distribution
FROM ubuntu:22.04

# Install Open3D system dependencies and pip
RUN apt-get update && apt-get install --no-install-recommends -y \
    libegl1 \
    libgl1 \
    libgomp1 \
    python3-pip \
    && rm -rf /var/lib/apt/lists/*

# Install Open3D from the PyPI repositories
RUN python3 -m pip install --no-cache-dir --upgrade pip && \
    python3 -m pip install --no-cache-dir --upgrade open3d

If you have an NVIDIA GPU and want to use it for computation (CUDA) or visualization, follow these directions.

To run GUI applications from the docker container, add these options to the docker run command line to ensure that docker has access to the:

GPU:

Intel (Mesa drivers): --device=/dev/dri:/dev/dri or --device=/dev/dri/card0:/dev/dri/card0 --device=/dev/dri/renderD128:/dev/dri/renderD128, depending on your hardware.

NVIDIA: --gpus 'all,"capabilities=compute,utility,graphics"'

No GPU (CPU rendering): CPU rendering is automaticaly selected if a GPU is not available.

X server: -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY

For example, the following commands will build and run the docker container with the Open3D viewer application, and provide access to the current directory as /root. Once the docker image is built, you can run the container from any folder that contains data you wish to visualize.

mkdir open3d-viewer-docker && cd open3d-viewer-docker
# Download Open3D viewer deb package.
wget https://github.com/isl-org/Open3D/releases/download/v0.18.0/open3d-app-0.18.0-Ubuntu.deb
# Build docker image in folder containing Open3D deb package.
docker build -t open3d-viewer -f- . <<EOF
FROM ubuntu:20.04
COPY open3d*.deb /root/
RUN apt-get update \
    && apt-get install --yes /root/open3d*.deb \
    && rm -rf /var/lib/apt/lists/*
ENTRYPOINT ["Open3D"]
EOF

# Allow local X11 connections
xhost local:root
# Run Open3D viewer docker image with the Intel GPU
docker run --device=/dev/dri:/dev/dri \
    -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY \
    -v "$PWD":/root open3d-viewer:latest
# Run Open3D viewer docker image with the NVIDIA GPU
docker run  --gpus 'all,"capabilities=compute,utility,graphics"' \
    -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY \
    -v "$PWD":/root open3d-viewer:latest
# Run Open3D viewer docker image without a GPU (CPU rendering)
docker run -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY \
    -v "$PWD":/root open3d-viewer:latest

Also see the docker tutorial for ROS for more information. Note that differences in hardware, OS drivers and OS packages may require you to modify these instructions.

Headless rendering#

If a GUI display server (X11 or Wayland) is not available (either in the docker container or the host OS), Open3D can still be used for headless rendering. In Ubuntu 20.04+ (with Mesa version 20.2+) this requires configuring the Mesa driver with an environment variable (EGL_PLATFORM=surfaceless):

mkdir open3d-headless-docker && cd open3d-headless-docker
wget https://raw.githubusercontent.com/isl-org/Open3D/v0.18.0/examples/python/visualization/render_to_image.py
# Build docker image
docker build -t open3d-headless -f- . <<EOF
FROM ubuntu:20.04
RUN apt-get update \
    && apt-get install --yes --no-install-recommends \
    libegl1 libgl1 libgomp1 python3-pip \
    && rm -rf /var/lib/apt/lists/*

# Install Open3D from the PyPI repositories
RUN python3 -m pip install --no-cache-dir --upgrade pip && \
    python3 -m pip install --no-cache-dir --upgrade open3d==0.18.0

# Configure Mesa EGL for headless rendering
ENV EGL_PLATFORM=surfaceless
WORKDIR /root/
ENTRYPOINT ["python3", "/root/render_to_image.py"]
EOF

# Run headless rendering example with Intel GPU
docker run --device=/dev/dri:/dev/dri \
    -v "$PWD":/root open3d-headless:latest
# Run headless rendering example with Nvidia GPU
docker run  --gpus 'all,"capabilities=compute,utility,graphics"' \
    -v "$PWD":/root open3d-headless:latest
# Run headless rendering example without GPU (CPU rendering)
docker run -v "$PWD":/root open3d-headless:latest

After running these commands, there will be two offscreen rendered images test.png and test2.png in the open3d-headless-docker folder.

Building Open3D in Docker#

If your current system does not support the minimum system requirements for building Open3D or if you have different versions of Open3D dependencies installed, you can build Open3D from source in docker without interfering with your system. This may be the case for older OS such as Ubuntu 16.04 or CentOS 7. We provide docker build scripts and dockerfiles to build Python wheels in various configurations. You can choose between different versions of Python, hardware architectures (AMD64, ARM64, CUDA) and developer vs release modes. Some sample configuration options available are shown below.

cd docker

./docker_build.sh cuda_wheel_py38_dev   # Python 3.8, AMD64, CUDA with MKL, developer mode
./docker_build.sh openblas-amd64-py310  # Python 3.10, AMD64 with OpenBLAS instead of MKL, release mode
./docker_build.sh openblas-arm64-py37   # Python 3.7, ARM64 with OpenBLAS, release mode

Run ./docker_build.sh without arguments to get a list of all available build configurations.

Note

You can control support for PyTorch and Tensorflow with environment variables: BUILD_PYTORCH_OPS=ON and BUILD_TENSORFLOW_OPS=ON