For the source code, please check the GitHub page. Pose estimation for a predefined CAD model using RGBD images. The CAD model shall be provided in standard PCD format (for reference, see here). The tool provides the following algorithms to estimate the pose:
The tool is fully implemented using ROS 2, however, a notebook is provided to help understanding the algorithms.
The tool provides an implementation for ROS 2 and notebooks to overview and understand the basics of the algorithms. For evaluating the ROS based version, please jump to ROS 2 Docker Image with GPU support. For evaluating and getting familiar with the algorithm in Jupyter notebooks, see notebooks documentation. For best practices regarding pose estimation, check best practices.
For ease of evaluation, you can build a Docker image right out-of-the-box. For using Docker, ensure, that Docker is installed.
For building the Docker image with tag pose_estimation, enter the repository, and run:
docker build -t pose_estimation .
After the image is built, run the image using the GPU as
docker run --rm -ti --gpus all -v <assets>:/asset --network host pose_estimation
The <assets> directory shall contain the configuration as a config.yaml file (see example) and the point cloud asset in PCD format, referenced by the configuration file.
Please note, that ROS2 communcation protocol DDS uses shared memory when nodes are executed on the same machine. To avoid this, we disable shared memory feature (see docker_trick.xml).
Subscriptions (to change, please see configuration):
/camera/color/image_raw/camera/aligned_depth_to_color/image_raw/camera/aligned_depth_to_color/camera_infoPublishers:
/pose_estimation/diag/image: image for diagnostics (sensor_msgs/Image)/pose_estimation/diag/points: point cloud for diagnostics (sensor_msgs/PointCloud2)/pose_estimation/pose: pose list in user defined format (pe_interface/msg/PoseList.msg), object2camera transformation (see src/pe_interface)If the diag is enabled in the configuration, the diagnostic topics will receive diagnostic images and point clouds after each successful detection. Please note, that the point cloud contains the original point cloud restored from the depth image and the fitted reference object (coloured with red). Note: using diagnostic messages considerably increases running time.