YOLOv5

YOLOv5 Object Detection for Pine Cone Images

• original video link: https://www.youtube.com/watch?v=T0DO1C8uYP8 (in Korean)
• brief instruction using Google Colaborator
  • so we can use free provided GPU (Runtime - Change runtime type - GPU)
• change directory to content
  • %cd /content/
• clone YOLOv5
  • !git clone https://github.com/ultralytics/yolov5.git
• install requirements
  • %cd /content/yolov5/
  • !pip install -r requirements.txt
• you can download weights 'best.pt' and 'last.pt', and jump to the detect section to use them from here
  • trained with more images and added 'best_ext.pt' and 'last_ext.pt'
• make a new folder and name it as 'dataset' and upload your dataset
  • uploaded files will be deleted after this session
  • you can upload your dataset to Google Drive, so you don't need to upload everytime
• split dataset into train and validation
  • only if the dataset hasn't been splitted. otherwise just set train and valid folders in data.yaml file
  • %cd /
  • from glob import glob
  • img_list = glob('/content/dataset/(location of images)/*.jpg')
  • from sklearn.model_selection import train_test_split
  • train, val = train_test_split(img_list, test_size=0.2, random_state=2000)
• save locations of train and val
  • with open('/content/dataset/train.txt', 'w') as f:
    • f.write('\n'.join(train) + 'n')
  • with open('/content/dataset/val.txt', 'w') as f:
    • f.write('\n'.join(val) + 'n')
• change data.yaml to indicate above train and val locations (can simply open the file and edit also)
  • import yaml
  • with open('/content/dataset/data.yaml', 'r') as f:
    • data = yaml.load(f)
  • data['train'] = '/content/dataset/train.txt' (or a train directory)
  • data['val'] = '/content/dataset/val.txt' (or a val directory)
  • with open('/content/dataset/data.yaml', 'w') as f:
    • yaml.dump(data, f)
  • modify nc as 1 since we only want to detect pincones
• training
  • %cd /content/yolov5/
  • !python train.py --img 640 --batch 16 --epochs 50 --data /content/dataset/data.yaml --cfg ./models/yolov5s.yaml --weights yolov5s.pt --name pinecones_yolov5s_results
• detecting
  • %cd /content/yolov5/
  • !python detect.py --weights /content/yolov5/runs/train/pinecones_yolov5s_results/weights/best.pt --img 640 --conf 0.5 --source (location of videos or photos)