Real-time pet detection and recognition with YOLO

• Download trained weight file from our release or train it by yourself.

• Clone YOLO Project

  git clone https://github.com/pjreddie/darknet 
  cd darknet 
  

• modify yolo source code

  vim examples/yolo.c
  

  Create a new array of labels string.

  char *pet_names[] = {"Abyssinian", "Bengal", "Birman", "Bombay", "British_Shorthair", "Egyptian_Mau", "Maine_Coon", "Persian", "Ragdoll", "Russian_Blue", "Siamese", "Sphynx", "american_bulldog", "american_pit_bull_terrier", "basset_hound", "beagle", "boxer", "chihuahua", "english_cocker_spaniel", "english_setter", "german_shorthaired", "great_pyrenees", "havanese", "japanese_chin", "keeshond", "leonberger", "miniature_pinscher", "newfoundland", "pomeranian", "pug", "saint_bernard", "samoyed", "scottish_terrier", "shiba_inu", "staffordshire_bull_terrier", "wheaten_terrier", "yorkshire_terrier"};
  

  Go to the test_yolo function and modify the input parameter of draw_detections.

  Usedraw_detections(im, dets, l.side*l.side*l.n, thresh, pet_names, alphabet, 37); replace it.

• Compile and test your yolo.

  Compile and link this project first.

  make clean
  make -j8
  

  Test

  ./darknet detector test cfg/pet.data cfg/yolov3-pet.cfg your-yolo-pet.weights ~/Your-dataset/JPEGImages/havanese_102.jpg 
  

  There will be an image named predictions.png in your darknet root directory.

  

If you want to train yolo by your self or train other dataset, you can refer to this guide.

• Clone this repo

  git clone https://github.com/ideaRunner/yolo-pet.git
  cd yolo-pet
  

• Download Pet Dataset

  The Oxford-IIIT Pet Dataset

  You can download the images and annotations by

  wget -c http://www.robots.ox.ac.uk/~vgg/data/pets/data/images.tar.gz
  wget -c http://www.robots.ox.ac.uk/~vgg/data/pets/data/annotations.tar.gz
  

  Unzip

  tar -xvf images.tar.gz
  tar -xvf annotations.tar.gz
  

  Change the directory name to which yolo need.

  mv images JPEGImages
  

• Clone YOLO Project

  git clone https://github.com/pjreddie/darknet
  cd darknet
  make
  

  Download Pretrained Convolutional Weights of Yolo-v3

  wget https://pjreddie.com/media/files/darknet53.conv.74

• Generate label files

  cd ..
  python pet_label generate
  

  Then there will be a Train_List.txt file and Val_List.txt file in the root directory and labels in labels directory. You must Keep labels and JPEGImages in same directory.

• Edit configure files

  vim darknet/cfg/pet.data
  

  copy, paste and save

  classes= 37
  train  = Your/Own/Path/Train_List.txt
  valid  = Your/Own/Path/Val_List.txt
  names = data/pet.names
  backup = pet_backup
  

  Don't forget to change the path as you generate before.

  vim darknet/data/pet.names
  

  copy, paste and save

  Abyssinian
  Bengal
  Birman
  Bombay
  British_Shorthair
  Egyptian_Mau
  Maine_Coon
  Persian
  Ragdoll
  Russian_Blue
  Siamese
  Sphynx
  american_bulldog
  american_pit_bull_terrier
  basset_hound
  beagle
  boxer
  chihuahua
  english_cocker_spaniel
  english_setter
  german_shorthaired
  great_pyrenees
  havanese
  japanese_chin
  keeshond
  leonberger
  miniature_pinscher
  newfoundland
  pomeranian
  pug
  saint_bernard
  samoyed
  scottish_terrier
  shiba_inu
  staffordshire_bull_terrier
  wheaten_terrier
  yorkshire_terrier
  

  Create yolov3.cfg and Edit the last serveal lines, change fliters to 126 and classes to 37. For yolov3, you need to find and change 3 times. For yolov2, you need to find and change 2 times. You also need to uncomment the train part batch and subdivisions in the first servel lines of this cfg file.

  You can use our .cfg file in this repo if you don't want to change by your self.

  mv darknet/cfg/yolov3-voc.cfg darknet/cfg/yolov3-pet.cfg 
  vim darknet/cfg/yolov3-pet.cfg 
  

  Make a directory for training backup. The name should be same as you defined in pet.data.

  mkdir darknet/pet_backups
  

• Train yolo

  cd darknet 
  ./darknet detector train cfg/pet.data cfg/yolov3-pet.cfg  darknet53.conv.74 
  

  The train result will be at your backup directory. You can fellow the guide above and test wheather it is good or not.