Modified 2017-12-11 by tanij
Modified 2017-12-09 by Igor Vasiljevic
This package contains a ROS node
dt_visual_odometry that produces monocular depth estimates on duckiebot. It also contains another node
apriltags_ros_center, which is slightly modified from
apriltags_ros to publish pixel locations, in order to benchmark the result on April tags. You need to have Tensorflow installed on your local machine.
To launch the node, clone the repository into
catkin workspace and download the Tensorflow model checkpoint file from Duckietown Dropbox to
checkpoint_dir. This Tensorflow model is obtained through taking a pretrained model on KITTI and further train it on Duckietown for 200K iterations with smoothness penalty set as 0.1. Use the command:
roslaunch dt_visual_odometry deepvo.launch ckpt_file:=checkpoint_dir robot_name:=robot_name
This publishes the depth heatmap the into
robot_name/VO/image/compressed as well as prints the predicted(from CNN) and actual(generated from April tags node, unit in meters) depth on any detected April tags. The scaling factor is calculated as the average predicted/actual depth for all detected April tags, and published under the topic
robot_name/VO/scale. To supress April tags detections for a higher refresh rate of depth heatmap, use
No questions found. You can ask a question on the website.