Setting up YoloV5
By Tom O’Donnell (tkodonne@purdue.edu)
You will need Python and Pip installed for this step. Make a new working directory for setting up Yolo.
git clone https://github.com/ultralytics/yolov5
Now installing all requirements for Yolo:
pip install -r yolov5/requirements.txt
(As a forewarning, we were unable to use Yolo on M1 Macs due to architecture issues. You may need to compile a few libraries from source to do so.)
Getting a dataset
To make a suitable dataset, you will need a large collection of images of the object you want to detect. The size is to be determined by you.
After collecting a set of images, you will need to add bounding boxes to the images and export the dataset into the YoloV5 format.
PRO Tip:
Use roboflow.com to simplify this process. Simply upload the images, and use the built-in bounding box feature to add bounding boxes.
You can also add preprocessing/augmentations to diversify your data as needed, however this step is not always necessary. (Then export the dataset in YoloV5 (pytorch) format).
Train YoloV5
Download the dataset you created, and paste all the files and folders into the root directory of your YoloV5 folder.
To train YoloV5, run the following command:
python train.py --img <image_size> --cfg yolov5s.yaml --batch 32 --epochs <epochs> --data data.yaml --weights yolov5s.pt --workers 24
This process will likely take forever, so please use a beefy PC. Also, increasing epochs will likely increase the accuracy of YoloV5, but it will increase training time and detection time hugely. Finally, these settings are not optimal for every use case. You will need to adjust these parameters.
Testing YoloV5
python detect.py --source 0for webcampython detect.py --source file.jpgfor image- and so on for other data types
Demo:
Detecting Enemy Robots via YOLOv5: https://youtu.be/4KRM93TuxyA