dsuchy / cvapr_yolo Goto Github PK

• Install required vesion of Python (3.8 or later).
• Clone this repository.

git clone https://github.com/syphzur/CVAPR_YOLO.git
cd ./CVAPR_YOLO

• Clone YOLOv5 repository.

git clone https://github.com/ultralytics/yolov5.git
cd ./yolov5

• Install required dependencies.

pip install -r requirements.txt

• To use transfer learning modify the train.py script provided in YOLOv5 repository to freeze all backbone layers.

freeze = ['model.%s.' % x for x in range(10)] 

• To train the network use modified script with proper path to dataset.yaml file.

python train.py --batch 48 --weights yolov5m.pt --data ../dataset.yaml --epochs 50 --cache --img 512

• To run inference use yolo_detector.py script that relies on script provided by Ultralytics. It can be used with variety of sources: webcam (0 as parameter), image, video, directory or even a YouTube video.

python yolo_detector.py 'https://youtu.be/NUsoVlDFqZg'
python yolo_detector.py 0 #webcam
python yolo_detector.py image.jpg
python yolo_detector.py directiory_with_images

• To run inference use without having YOLOv5 repository cloned use std_detector.py script created with Pytorch. It can be used with image path passed as argument.

python std_detector.py image.jpg

Dataset was downloaded from Kaggle and modified to YOLOv5 format.