sliding_window_object_detector_trainer_node

What is this?

Node to train jsk_perception/SlidingWindowObjectDetector using binary support vector machine.

The object is assigned a label of +1 and -1 otherwise. The SVM used is from the OpenCV Library with default set to RBF Kernel and 10-Fold Cross Validations.

Parameters

  • ~dataset_path (string, required)

    Folder name where ~object_dataset_filename and ~nonobject_dataset_filename resides.

    It should end with /.

  • ~object_dataset_filename (string, required)

    Rosbag file name of the object (positive) training set.

    The bag file must contain ~object_dataset_topic topic.

  • ~object_dataset_topic (string, default: /dataset/roi)

    Topic name of sensor_msgs/Image which is a set of positive training examples.

  • ~nonobject_dataset_filename (string, required)

    Rosbag file name of the non-object (negative) training set.

    The bag file must contain ~nonobject_dataset_topic topic.

  • ~nonobject_dataset_topic (string, default: /dataset/background/roi)

    Topic name of sensor_msgs/Image which is a set of negative training examples.

  • ~classifier_name (string, required)

    Path to trained SVM classifier output file.

    .xml or .yaml format is supported.

  • ~manifest_filename (string, default: sliding_window_trainer_manifest.xml)

    Path to manifest file which contains parameters of the trainer such as trainer window size, save directory, etc.

    .xml or .yaml format is supported.

    ../../_images/trainer_manifest.png

  • ~swindow_x (int, required)

  • ~swindow_y (int, required)

    Images in training dataset are resized to this size (width, height) before training SVM.

Sample

roslaunch jsk_perception sample_sliding_window_object_detector_trainer.launch

and wait a few minutes until “Trained Successfully” message appears.