/Train-YOLOv3-Custom-Object-Detector-with-Darknet

Primary LanguagePython

Train YOLOv3 Custom Object Detector with Darknet

Link: Google Image Dataset API

You can follow this video too!

Link: Custom Object Detecto

First we have to collect and manage dataset

Link: Google Image Datasets API

  1. Download or clone Train-YOLOv3-Custom-Object-Detector-with-Darknet repository Link

pip install -r requirements.txt

  1. Open command prompt from the directory where you've donwloaded/cloned the repository

  2. Run on the cmd: python main.py downloader --classes Aircraft Weapon --type_csv train --limit 1000 --multiclasses 1 It will download dataset consist of classes that we want.

Explaination

  • --classes Aircraft Weapon:Aircraft and Weapon are classes that we want to download
  • --type_csv train : We want training dataset
  • --limit 1000 : not to download more than 1000 images(For each class)
  • --multiclasses 1: We want our classes to be in one folder

Then data will be downloaded in OID/Dataset/train folder

  1. Change the class.txt according to the classes of the dataset.For aour case, it was
    Aircraft
    Weapon

  2. Run python convert_annotations.py . It will create an extra annotation file after each of the image file

  3. Copy all data from Aircraft_Weapon folderr and paste them into C:\darknet-master\data\obj folder

Our dataset creating step is complete

Training object detection based on our dataset:

  1. Go config folder in darknet and copy yolov3.cfg and rename the copied file to yolov3-custom.cfg
  2. Edit yolov3-custom.cfg Editing part:
  • Make comment lines in Testing(#batch=1,#subdivisions=1)
  • Uncomment lines in training(batch=64,subdivisions=32) Here subdivisions 16>32>64>128. Chooese depending on your pc's hardware
  • max_batches = 4000 (classes*2000)
  • steps=3200,3600 (80%,90% of our max_batches)
  • Press ctrl+f and find yolo.You'll find 3 yolo layers.Change the classes=2 in 3 yolo layer and filters=21 ((classes+5)*3) in [convolutional] layer above each yolo layers -save

  1. GO to the data folders and make a file named obj.names and edit and save them accordingly(We've changes Airchraft and weapon class to Car and food classes.You have to change accorging you classes)
    Car
    Food

  2. Then make a file named obj.data and edit and save them accordingly
    classes = 2
    train = C:/darknet-master/data/train.txt
    validation = C:/darknet-master/data/test.txt
    names = C:/darknet-master/data/obj.names
    backup = C:/darknet-master/backup/

  3. Make a backup file in the darknet directory

  4. Donwload generate_train.py and put it on darknet directory

  5. Open cmd from darknet directory and run command: python generate_train.py (It will create a file path of every singe dataset images in data folder names as train.txt)

  6. Then go to YOLO: Real-Time Object Detection and download the file pre-trained weight file

  7. It will download a file named: darknet53.conv.74 , you have to pase it to: C:\darknet-master\build\darknet\x64

  8. Finally run cmd in location: C:\darknet-master\build\darknet\x64 and run a command:
    darknet.exe detector train C:/darknet-master/data/obj.data C:/darknet-master/cfg/yolov3-custom.cfg darknet53.conv.74

It'll start training and best weight will be saved in C:/darknet-master/backup/yolov3-custom_last.weights

If somehow training stops in the mid, then if you run the training again.It won't strart from begining as we've weights in our backup file

Few commands:

  • If you want to train from the beginning, then use flag in the end of training command: -clear

To test our model on image:

  • darknet.exe detector test C:/darknet-master/data/obj.data C:/darknet-master/cfg/yolov3-custom.cfg C:/darknet-master/backup/yolov3-custom_last.weights 1.jpg

To test our model in video: -darknet.exe detector demo C:/darknet-master/data/obj.data C:/darknet-master/cfg/yolov3-custom.cfg C:/darknet-master/backup/yolov3-custom_last.weights 1.mp4