Azure Kinect with Open3D

Azure Kinect is only officially supported on Windows and Ubuntu 18.04.

Installation

Install the Azure Kinect SDK

Follow this guide to install the Azure Kinect SDK (K4A).

On Ubuntu, you’ll need to set up a udev rule to use the Kinect camera without sudo. Follow these instructions.

After installation, you may run k4aviewer from the Linux terminal or k4aviewer.exe on Windows to make sure that the device is working.

Currently, Open3D supports the Azure Kinect SDK version v1.2.0, though future versions might also be compatible.

If you’re using Ubuntu 16.04, the Azure Kinect SDK is not officially supported. We have an unofficial workaround, see Unofficial Ubuntu 16.04 workaround for details.

Using Open3D from Pip or Conda

If you’re using Open3D installed via Pip or Conda, Open3D’s Azure Kinect features should work out-of-the box if K4A is installed in the system in the recommended way. Open3D will try to load the K4A dynamic library automatically at runtime, when a K4A related feature within Open3D is used.

On Ubuntu, the default search path follows the Linux convention.

On Windows, Open3D will try to load the shared library from the default installation path. For example, for K4A v1.2.0, the default path is C:\Program Files\Azure Kinect SDK v1.2.0. If this doesn’t work, copy depthengine_x_x.dll, k4a.dll and k4arecord.dll to where Open3D Python module is installed if you’re using Open3D with Python, or to the same directory as your C++ executable.

You can get Open3D’s Python module path with the following command:

python -c "import open3d as o3d; import os; print(os.path.dirname(o3d.__file__))"

Compile from Source

To build Open3D from source with K4A support, set BUILD_AZURE_KINECT=ON at CMake config step. That is,

cmake -DBUILD_AZURE_KINECT=ON -DOTHER_FLAGS ..

Open3D Azure Kinect Viewer

Open3D Azure Kinect Viewer is used for previewing RGB and depth image stream captured by the Azure Kinect sensor.

Open3D provides Python and C++ example code of Azure Kinect viewer. Please see examples/cpp/AzureKinectViewer.cpp and examples/python/ReconstructionSystem/sensors/azure_kinect_viewer.py for details.

We’ll use the Python version as an example.

python examples/python/ReconstructionSystem/sensors/azure_kinect_viewer.py --align_depth_to_color
azure_kinect_viewer_aligned.png

When recording at a higher resolution at a high framerate, sometimes it is helpful to use the raw depth image without transformation to reduce computation.

python examples/python/ReconstructionSystem/sensors/azure_kinect_viewer.py
azure_kinect_viewer_unaligned.png

When the visualizer window is active, press ESC to quit the viewer.

You may also specify the sensor config with a json file.

python examples/python/ReconstructionSystem/sensors/azure_kinect_viewer.py --config config.json

An sensor config will look like the following. For the full list of available configs, refer to this file.

{
    "camera_fps" : "K4A_FRAMES_PER_SECOND_30",
    "color_format" : "K4A_IMAGE_FORMAT_COLOR_MJPG",
    "color_resolution" : "K4A_COLOR_RESOLUTION_720P",
    "depth_delay_off_color_usec" : "0",
    "depth_mode" : "K4A_DEPTH_MODE_WFOV_2X2BINNED",
    "disable_streaming_indicator" : "false",
    "subordinate_delay_off_master_usec" : "0",
    "synchronized_images_only" : "false",
    "wired_sync_mode" : "K4A_WIRED_SYNC_MODE_STANDALONE"
}

Open3D Azure Kinect Recorder

Open3D Azure Kinect Viewer is used for recording RGB and depth image stream to an MKV file.

Open3D provides Python and C++ example code of Azure Kinect viewer. Please see examples/cpp/AzureKinectRecord.cpp and examples/python/ReconstructionSystem/sensors/azure_kinect_recorder.py for details.

We’ll use the Python version as an example.

python examples/python/ReconstructionSystem/sensors/azure_kinect_recorder.py --output record.mkv

You may optionally specify the camera config when running the recorder script.

When the visualizer window is active, press SPACE to start or pause the recording or press ESC to quit the recorder.

azure_kinect_recorder.png

Open3D Azure Kinect MKV Reader

The recorded MKV file uses K4A’s custom format which contains both RGB and depth information. The regular video player may only support playing back the color channel or not supporting the format at all. To view the customized MKV file, use the Open3D Azure Kinect MKV Reader.

Open3D provides Python and C++ example code of Open3D Azure Kinect MKV Reader. Please see examples/cpp/AzureKinectMKVReader.cpp and examples/python/ReconstructionSystem/sensors/azure_kinect_mkv_reader.py for details.

python examples/python/ReconstructionSystem/sensors/azure_kinect_mkv_reader.py --input record.mkv
azure_kinect_mkv_reader.png

Note that even though the recorder records the unaligned raw depth image, the reader can correctly wrap the depth image to align with the color image.

To convert the MKV video to color and depth image frames, specify the --output flag.

python examples/python/ReconstructionSystem/sensors/azure_kinect_mkv_reader.py --input record.mkv --output frames
azure_kinect_mkv_reader_extract.png

Unofficial Ubuntu 16.04 workaround

The Azure Kinect SDK is not officially supported on Ubuntu 16.04. We provide unofficial support for experimental purposes.

Using Open3D Python packages

For Python Open3D, run

pip install open3d_azure_kinect_ubuntu1604_fix

The open3d_azure_kinect_ubuntu1604_fix package contains 4 shared libs:

  • libstdc++.so: copied from ubuntu 18.04, needed by libdepthengine.so

  • libdepthengine.so: copied from K4A installation on ubuntu 18.04

  • libk4a.so: build from source on ubuntu 16.04

  • libk4arecord.so: build from source on ubuntu 16.04

The open3d_azure_kinect_ubuntu1604_fix will preload the shared libs and set LD_LIBRARY_PATH which are then used by dlopen when the Kinect library is loaded from the compiled module.

After installing open3d_azure_kinect_ubuntu1604_fix, import Open3D as usual with import open3d. Open3D will try to load the shared libraries at initialization time in __init__.py.

Compiling Open3D from source

First, install open3d_azure_kinect_ubuntu1604_fix package in your Python environment. This package contains the headers that are required for compiling Open3D with Azure Kinect support.

When building Open3D from source, set the flag -DBUILD_AZURE_KINECT=ON at CMake configure time. CMake will then try to detect the location of the open3d_azure_kinect_ubuntu1604_fix package using the Python executable available from the current PATH. Therefore, when running CMake, make sure that the same Python environment where open3d_azure_kinect_ubuntu1604_fix was installed is activated.

If you build a C++ binary, you’ll still need to ensure that LD_LIBRARY_PATH contains the directory which contains libstdc++.so and libdepthengine.so at runtime. If you build the compiled Open3D Python module, Open3D’s __init__.py will try to import open3d_azure_kinect_ubuntu1604_fix to append the LD_LIBRARY_PATH automatically.