Overview

This is a common interface for registering point clouds captured using a depth camera (Kinect V2 or Azure Kinect), supporting COLMAP and SLAM-based solutions. A general data capturing program is included.

Binaries

• rgbdrec: Basic RGB-D sequence recorder
• rgbdreg-colmap, rgbdreg-orbslam2: Point cloud registration (via odometry/SFM): supports COLMAP and ORB_SLAM2. Uses data format output by rgbdrec
  • rgbdreg-colmap: registration method based on COLMAP. Note the SFM is performed using color images only and is initially in an arbitrary unit; we transform all matched feature points to camera space and compare the depth to the depth image to correct the scale. Requires colmap to be in system PATH.
  • rgbdreg-orbslam2: registration method based on ORB_SLAM2. This method actually uses RGB-D information of each frame and produces correctly scaled camera poses.
• rgbdreg-viewer: Visualize camera pose estimation. After running one of the above, use rgbdreg-viewer <data-folder> to ensure result is good.

Format

Each registration program produce a file poses.txt (readable with numpy.loadtxt) in the data directory. This is a text file of floats, separated by spaces and newlines, of shape [N, 12] where N is the number of images; each row can be viewed as a [3, 4] row-major rigid-body transform matrix (bottom row omitted) which transforms a point in homogeneous coordinates from camera space to world space. As usual, the 3 leftmost columns are the rotation and right column is the translation.

Regardless of program used, the output coordinate system will be right-handed with axis orientations [x, y, z] = [right, up, backward (screen to your face)].

See rgbdreg-viewer.cpp for an example program reading poses.txt.

Setup

Dependencies

• OpenCV 3+
• Eigen 3
• Boost (really I only use program_options, filesystem, process)

Optional Dependencies

• For recording, at least one of
  • libk4a (Azure Kinect SDK)
  • libfreenect2
• Registration software
  • For COLMAP just build and install https://github.com/colmap/colmap.
  • For ORB_SLAM2 the setup is more annoying. I use the binary ORB vocab format introduced by poine: https://github.com/poine/ORB_SLAM2. You can use my fork: https://github.com/sxyu/ORB_SLAM2 or one of many other versions to support OpenCV4 and make the build easier. Set set ORB_SLAM2_ROOT_DIR environment variable to the repo root.
• Visualization
  • rgbdreg-viewer needs my OpenGL viewer https://github.com/sxyu/meshview. Follow instructions there to install.

Build

Build using cmake as usual: mkdir build && cd build && cmake .. && make -j12

Credit

Out of laziness I reuse a lot of code from my other project https://github.com/sxyu/avatar, which also borrows from https://github.com/augcog/OpenARK