ROS package for official YOLOv7

Note
This ROS noetic package is an implementation of the official yolov7, most credit goes to them.

Fig.1 RViz view of detection projection from a Gazebo Simulation

🚀 Installation from source

First you need to download all required source, like the ros-image-transport-py package, that manage automaticly image stream. Moreover you might need some python package like OpenCV, you can install them by simply execute pip install opencv-python or better compile from source for GPU.

⚙️ Clone and Compile

Clone thoose two packages inside the src directory of your catkin workspace.

git clone https://github.com/alexandrefch/ros-yolov7.git
git clone https://github.com/alexandrefch/ros-image-transport-py.git

Now you can compile them using your catkin tool (catkin_make, catkin_build ...).

📋 Configuration

Weight

You can download some of the the YOLOv7 weights from the official repository or use your own.

Classes

Classes should be stored inside a simple text file where each single line represent a classe, like in the exemple below.

car
bus
person
bike
motorcycle
traffic light
truck
person

Launch file

<?xml version="1.0"?>
<launch>
    <node pkg="yolov7" type="YoloV7.py" name="yolov7">
        <!-- Path to your weight -->
        <param name="weights_path" type="str" value="path_to_your/weights/weight.pt"/>
        <!-- Path to a class_labels.txt file, if you leave it empty then no class labels are visualized.-->
        <param name="classes_path" type="str" value="path_to_your/class_labels/labels.txt" />
        <!-- Input image topic name to subscribe to -->
        <param name="img_topic" type="str" value="/camera/image" />
        <!-- [optional]  Confidence threshold (default=0.25) -->
        <param name="conf_thresh" type="double" value="0.20" />
        <!-- [optional]  Intersection over union threshold (default=0.45) -->
        <param name="iou_thresh" type="double" value="0.45" />
        <!-- [optional]  Queue size for publishing (default=3) -->
        <param name="queue_size" type="int" value="1" />
        <!-- [optional] Image size to which to resize each input image before feeding into the network (the final output is rescaled to the original image size) (default=640) -->
        <param name="img_size" type="int" value="640" />
        <!-- [optional] Flag whether to also publish image with the visualized detections (default=false) -->
        <param name="visualize" type="bool" value="true" />
        <!-- [optional] Torch device 'cuda' or 'cpu' (default="cuda") -->
        <param name="device" type="str" value="cuda" />
        <!-- [optional] Node frequency (default=10) -->
        <param name="frequency" type="int" value="10" />
    </node>
</launch>

✒️ Citation

If you use this package in a scientific way or public way, don't forget to mention the official Yolov7 paper, and star their work on their github repo.

@article{wang2022yolov7,
  title={{YOLOv7}: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors},
  author={Wang, Chien-Yao and Bochkovskiy, Alexey and Liao, Hong-Yuan Mark},
  journal={arXiv preprint arXiv:2207.02696},
  year={2022}
}