Upgraded version here : https://github.com/proplayer2020/annot-tool-GUI
Have you ever wanted to train your custom image detection model ?
This repository will help annotate data easily for that.
Use with this repo : https://github.com/AlexeyAB/darknet
open a terminal and type following commands :
git clone https://github.com/proplayer2020/annot_data/
cd annot_data
Then, go into "images" folder and create as many folders as the number of classes (starting from 0) and put your images here (class order isnt important just do not shuffle classes into the same folder).
Here is an example (if the model will have to recognize flowers and cars):
images/0/* (car images here)
images/1/* (flower images here)
Just run this command :
python3 annot_tool.py 0
Organize all classes in directories (one directory per class), for example :
images/0/*
images/1/*
images/2/*
python3 annot_tool.py 0 # for images/0/*
python3 annot_tool.py 1 # for images/1/*
python3 annot_tool.py 2 # for images/2/*
It is possible to change names later while training to change "obj.names" with all your names, in this order.
Just click on the first position IN TOP LEFT CORNER OF THE OBJECT, and then click in THE BOTTOM RIGHT CORNER, you will see a red bouding box that isn't moving with your cursor (if the bouding box moves, click again in the right position), then, just close the window, another image will appear.
Useful keybind in the annotation program :
- 'q' to quit
- 'n' to skip image
- 'c' to crop (for zooming on a restricted area, this cannot be undoned)
When all images are annoted, copy paste them in the images folder (so delete all 0,1,2 folders) Then enter in terminal
python3 traintest_split.py --nb_test <number of image to put in test set> --path <path to image folder (containing all images with all .txt)> --nb_class <number of classes in the dataset>
This code will split all annoted images into two .txt files : train.txt and test.txt. Images_nb arg is to define the number of images in test.txt set (at least 10-15 percent) All other images will go into train set.
The train command will use those files (where has been put all images name)
Use the 'see_data.py' script to see if all images were annoted correctly
In the terminal :
python3 see_data.py --datapath <here the path of the train.txt or test.txt created previously with traintest_split.py>
If some annotations are bad, go into the console (e.g the terminal) and delete image, txt file and the image name in the train.txt (or test.txt), else, the model will ignore the image or throw an error.
Download this repository following instructions : https://github.com/AlexeyAB/darknet
Then compile it to make it work : https://github.com/AlexeyAB/darknet#how-to-compile-on-linux-using-make
Take the two files generated in "result" folder in this repository (train.txt and test.txt) and put them in the data folder in the darknet repository
Prepare for training (YOLO) : https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects Prepare for training (YOLO tiny) : https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects
NOTE : if "darknet.exe" is written in the command, replace it with ./darknet on linux.
✔️ Easy to use
✔️ Totally free to use and no liscence
✔️ 99% working lol
✔️ Im active and will respond to all of your concerns
❌ A bit slow to refresh (made with matplotlib lol)
❌ Sometimes the bounding box do weird things, (in that case type 'n' to go to next image and then delete the faulty image)
Now you are all set to train a custom model with the github below !
feel free to post issue, im active and i will try to respond to your questions