ROS package for official YOLOv7

This repo contains a ROS noetic package for the official YOLOv7. It wraps the official implementation into a ROS node (so most credit goes to the YOLOv7 creators).

Note

There are currently two YOLOv7 variants out there. This repo contains the implementation from the paper YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors.

Requirements & Getting Started

Following ROS packages are required:

• vision_msgs
• geometry_msgs

First, clone the repo into your catkin workspace and build the package:

git clone https://github.com/lukazso/yolov7-ros.git ~/catkin_ws/src/
cd ~/catkin_ws
catkin build yolov7_ros

The Python requirements are listed in the requirements.txt. You can simply install them as

pip install -r requirements.txt

Download the YOLOv7 weights from the official repository.

The package has been tested under Ubuntu 20.04 and Python 3.8.10.

Usage

Before you launch the node, adjust the parameters in the launch file. For example, you need to set the path to your YOLOv7 weights and the image topic to which this node should listen to. The launch file also contains a description for each parameter.

roslaunch yolov7_ros yolov7.launch

Each time a new image is received it is then fed into YOLOv7.

Notes

• The detections are published using the vision_msgs/Detection2DArray message type.
• The detections will be published under /yolov7/out_topic defined in the launch file.
• If you set the visualize parameter to true, the detections will be drawn into the image, which is then published under /yolov7/out_topic/visualization.

MIL Training (TODO: Make this setup process more automated.)

• Go to your labelbox project and export the labels.
• Outside of the repository, clone this repo: git@github.com:ultralytics/JSON2YOLO.git.
• Use it to convert your labelbox export into a directory containing the images and corresponding labels.
• This can be accomplished by moving your labelbox export within this repo, change the filename within labelbox_json2yolo.py, run this python script. (This can take a while if your dataset is large.)
• This will create a directory with the same name as your .json file you exported.
• Within this new directory, you will see an images folder and a labels folder.
• Create a train and val folder within both the images and labels folders, and place a certain number of files within each folder. (Make sure the files correspond with each other. Ex. if image1.png is in train, ensure image1.txt is in labels.)
• Within the file set as data.yaml in the launch file, make the train go to the path <zip file name>/images/train, and make the val go to path <zip file name>/images/val
• Move this the folder containing the images and labels to the src directory of the yolov7_ros package.
• Move the .yaml within the folder you just moved into yolov7-ros/src/data.
• Confirm in cfg/training/yolov7.yaml the number of possible classes.
• Run the following command, feel free to change any parameters as you see fit. You can see what params you can change within train.py. python3 train.py --workers 8 --device 0 --batch-size 4 --data data/buoys.yaml --img 640 360 --cfg cfg/training/yolov7.yaml --weights 'yolov7_training.pt' --name yolov7 --hyp data/hyp.scratch.p5.yaml
• The resulting model (weights) will appear in "yolov7-ros/src/runs/train/yolov71i/weights/best.pt" where that i is the number of iteration. The latest training results should have the bigger i, but double check the date just in case.

MIL Single Image Testing

python3 detect.py --weights runs/train/<path to best.pt> --conf 0.7 --img-size 640 --source buoys/images/<path to any image>