Perception training

Get your set-up ready

• Get two people
• Take the kinect with the camera stand
• Take a computer with the ros-kinetic-freenect-launch and ros-kinetic-image-recognition packages installed (sudo apt install ros-kinetic-freenect-launch ros-kinetic-image-recognition)
• If you had to install the packages, execute 'rospack profile' to reindex the package list
• Get the object
• Get a table to put the objects on
• Get the turning table

Annotate objects

• Start the freenect driver: roslaunch freenect_launch freenect.launch
• Start the annotation GUI: rosrun image_recognition_rqt annotation_gui
• Annotate some images (minimum of 20 per object) - Use the turning table to get images from different perspective

More info: https://github.com/tue-robotics/image_recognition

Now train the neural net

• Start the GUI: rosrun image_recognition_tensorflow_rqt train_gui

More info: https://github.com/tue-robotics/image_recognition and https://github.com/tue-robotics/image_recognition/tree/master/image_recognition_tensorflow_rqt

• Use 1000 steps and 100 batch or ask @Rayman (Ramon Wijnands)

Test the perception

• Start the GUI rosrun image_recognition_rqt test_gui

More info: https://github.com/tue-robotics/image_recognition and https://github.com/tue-robotics/image_recognition/tree/master/image_recognition_rqt

Put everything in the right directory

Annotated verified images

~/MEGA/data/$(arg robot_env)/training_data/annotated/verified

Tensorflow output graph and labels

~/MEGA/data/$(arg robot_env)/models/tensorflow_ros/output_labels.txt ~/MEGA/data/$(arg robot_env)/models/tensorflow_ros/output_graph.pb