idearunner / chongaiyoujia

An app for pets, implemented pet detection and recognition.

Android app for real-time pet detection.

This project is developed by TensorFlow android example.

Requisite

Android 5.0 (API 21) or higher is required.

Downloading and Running App

• Install Android Studio from android.com

• Clone this repo.

  git clone https://github.com/ideaRunner/chongaiyoujia.git

• Download trained neural network model from release.

  cd assets
  wget https://github.com/ideaRunner/chongaiyoujia/releases/download/1.0.0/yolov2-tiny-pet_40000.pb
  

• Open android studio, open this project, follow it's instruction to download those libraries you need. (If you are in Chinese, you will need to setup proxy to download from google)

• Build (or Rebuild) project.

• Connect your mobile phone, make sure you have turned on USB debugging mode.

• Click the green run button on Android Studio, find your connected device and click OK.

• Wait one minute and the app could be installed on your phone, you can run it to detect 37 classes pets.

Pet Breeds

12 cat and 25 dog breeds from Oxford-IIIT Pet Dataset

• Cat

  Abyssinian
  Bengal
  Birman
  Bombay
  British_Shorthair
  Egyptian_Mau
  Maine_Coon
  Persian
  Ragdoll
  Russian_Blue
  Siamese
  Sphynx
  

• Dog

  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
  

Generate Your Own App

Train YOLO

For how to train yolo to detect pets or detect your own objects, follow this page.

Convert YOLO model to Tensorflow model

• Clone and install DarkFlow.

• Convert. ./flow --model cfg/your-tiny-yolo.cfg --load bin/your-tiny-yolo.weights --savepb --verbalise

  After running the command two files will appear in the ./built_graph directory:

  your-tiny-yolo.meta;
  your-tiny-yolo.pb;
  

Implement trained model

• Modify DetectorActivity

  private static final String YOLO_MODEL_FILE = "file:///android_asset/your-yolo-model.pb"; private static final DetectorMode MODE = DetectorMode.YOLO;

• Modify TensorFlowYoloDetector

  Change NUM_CLASSES and LABELS to what you have trained.

Run

After completing the above, you can run and download to your mobile device.

Quick Troubleshooting

Q1: Convert model by darflow AssertionError:

AssertionError: expect 63820056 bytes, found 63820060

Modify the line self.offset = 16 in the ./darkflow/utils/loader.py file and replace with self.offset = 20.