Image matching

System

• ROS2 Humble
• Pytorch
• Numpy
• OpenCV
• CUDA

Quickstart

In one terminal:

source install/setup.bash
ros2 run image_matching simple_matcher_node

In another terminal:

source install/setup.bash
ros2 service call /image_matching/toggle_template bb_perception_msgs/srv/IMPoseEstimatorToggleTemplate "template_name: 'Task04_Tagging_01.png'
camera_frame_id: 'auv4/front_cam_optical'
enable: true"

See image_matching/image for visualization. Point correspondences are published at image_matching/point_correspondences.

Notes

To obtain a pose from the output, run PnP on the 2D–3D correspondences (e.g., see pose_estimator/points_pose_estimator_node). We don't run PnP here since these raw correspondences can be combined with other detections downstream for a better estimate.

Use XFeat for general, upright camera matching. For matching between images with large orientation differences (e.g., drone imagery), try SIFT-FLANN or DALF.

See this commit for the old image matchers that the seniors used.

References

(Outdated) https://github.com/Shiaoming/Python-VO