This package provides a ROS wrapper for YOLOv3 based on PyTorch-YOLOv3. For consistency, the messages are based on the darknet_ros package. The package has been tested with Ubuntu 16.04 and ROS Kinetic on a NVIDIA Jetson TX2.

Authors: Vasileios Vasilopoulos ([email protected]), Georgios Pavlakos ([email protected])

Affiliation: University of Pennsylvania, GRASP Lab

To download the prerequisites for this package (except for ROS itself), use the ROS-default rosdep to install the required dependencies:

cd /your/catkin_ws
rosdep install --from-paths src --ignore-src -y

Alternatively you can use the python standard requirements.txt file to install the dependencies as well. Navigate to the package folder and run:

$ sudo pip install -r requirements.txt

Aftre making sure the required dependencies are installed following the instructions above, navigate to your catkin workspace and run:

$ catkin_make --pkg yolov3_pytorch_ros

1. First, make sure to put your weights in the models folder. This should automatically be done for the default configuration during the compilation process, however if you'd like to use your own make sure that they exist before you use them. For the training process in order to use custom objects, please refer to the original YOLO page.

By default during the build the following pre-trained weights are downloaded:

# download yolov3.weights
wget http://pjreddie.com/media/files/yolov3.weights
# dowload yolov3-tiny.weights
wget http://pjreddie.com/media/files/yolov3-tiny.weights

2. The default settings (using yolov3.weights) in the launch/detector.launch file should work, all you should have to do is change the image topic you would like to subscribe to:

roslaunch yolov3_pytorch_ros detector.launch image_topic:=/your/image/topic

You can also try out the yolov3-tiny.weights by simply passing in different arguments at launch. This is the recommended usage on a CPU for dramatically increased framerate with slightly reduced performance:

roslaunch yolov3_pytorch_ros detector.launch image_topic:=/your/image/topic config_name:=yolov3-tiny.cfg weights_name:=yolov3-tiny.weights confidence:=0.1

The confidence argument can be adjusted to set a threshold for detected objects to be ignored.

Alternatively you can modify the parameters in the launch file, recompile and launch it that way so that no arguments need to be passed at runtime.

• image_topic (string)

  Subscribed camera topic.

• weights_name (string)

  Weights to be used from the models folder.

• config_name (string)

  The name of the configuration file in the config folder. Use yolov3.cfg for YOLOv3, yolov3-tiny.cfg for tiny YOLOv3, and yolov3-voc.cfg for YOLOv3-VOC.

• classes_name (string)

  The name of the file for the detected classes in the classes folder. Use coco.names for COCO, and voc.names for VOC.

• publish_image (bool)

  Set to true to get the camera image along with the detected bounding boxes, or false otherwise.

• detected_objects_topic (string)

  Published topic with the detected bounding boxes.

• detections_image_topic (string)

  Published topic with the detected bounding boxes on top of the image.

• confidence (float)

  Confidence threshold for detected objects.

• image_topic (sensor_msgs::Image)

  Subscribed camera topic.

• detected_objects_topic (yolov3_pytorch_ros::BoundingBoxes)

  Published topic with the detected bounding boxes.

• detections_image_topic (sensor_msgs::Image)

  Published topic with the detected bounding boxes on top of the image (only published if publish_image is set to true).

The YOLO methods used in this software are described in the paper: You Only Look Once: Unified, Real-Time Object Detection.

If you are using this package, please add the following citation to your publication:

@misc{vasilopoulos_pavlakos_yolov3ros_2019,
  author = {Vasileios Vasilopoulos and Georgios Pavlakos},
  title = {{yolov3_pytorch_ros}: Object Detection for {ROS} using {PyTorch}},
  howpublished = {\url{https://github.com/vvasilo/yolov3_pytorch_ros}},
  year = {2019},
}