Image classification, object detection and image classification using neural networks

This Darknet fork is based mainly on AlexeyAB's Darknet fork, which is the best maintained and containing Yolo V2-3-4 detection networks, after pjreddie's original Darknet implementation (mainly for classification) and completed with ArtyZe's code for image segmentation.

For more information about the useage, please refer to the readme files of the respective forks.

Main features

• Image classification using ~20 state of the art networks (https://pjreddie.com/darknet/imagenet/)
• Object detection using Yolo networks (https://pjreddie.com/darknet/yolo/)
• Image segmentation using unet networks (https://github.com/ArtyZe/yolo_segmentation/blob/master/README.md)
• Text generation using rnns, GAN nightmare, etc...

Building

Recommended prerequisites

• CMake 3.18+
• OpenCV 2.4+
• CUDA 10.2+
• cuDNN 8.0.2+

Building

Linux/MacOS

Using CMake:

git clone https://github.com/kbarni/darknet
cd darknet
mkdir build_release
cd build_release
cmake ..
cmake --build . --target install --parallel 8

Using make: edit the options at the beginning of the provided Makefile, then run make.

Windows

• Clone this repository
• Run CMake (gui)
• Select the source and build folders, run Configure and Generate
• Open the created file with Visual studio, then Build solution.

Useage

Classification

• Create a file containing the class names (labels.txt), the training images (training.txt) and validation images (validation.txt). The image names must contain the class name (as defined in lables.txt).
• Create a data file containing:

classes = [n]
train = train.txt
valid = valid.txt
labels = labels.txt
backup = backup/

• Create a config file or choose one from the cfg folder

• Train the model:

  ./darknet classifier train mydata.data network.cfg

• Test the trained network on an image:

  ./darknet classifier test mydata.data network.cfg network.weights testimage.jpg

For a more detailed description check the CIFAR-10 classifier training tutorial.

Detection

• Generate label file for your files in format (x,y,w,h are relative to the image size):

• Create a data file as above, using a yolo network

• Train the model:

  ./darknet detector train mydata.data network.cfg

• Test the trained network on an image:

  ./darknet detector test mydata.data network.cfg network.weights testimage.jpg

For more details, refer to the Yolo tutorial and AlexeyAB's wiki.

Segmentation

• Create a file containing the class names (names.list), the training images (training.txt) and image labels (labels.txt). The label images must have the same name than the training images; they have to be 8 bit PNG images with every gray level corresponding to a class.
• Create a data file containing:

classes = [n]
train = train.txt
labels = labels.txt
backup = backup/

• Choose a unet network from the cfg folder

• Train the model:

  ./darknet segmenter train mydata.data network.cfg

• Test the trained network on an image:

  ./darknet segmenter test mydata.data network.cfg network.weights testimage.jpg

C++

For using Darknet in C++ with OpenCV, please refer to the sample files in the examples folder.

Further reading:

• Joseph Redmon's (pjreddie) darknet homepage: https://pjreddie.com/darknet/
• Alexey Bochkovskiy's (AlexeyAB) darknet wiki: https://github.com/AlexeyAB/darknet/wiki
• AlexeyAB's darknet readme: https://github.com/AlexeyAB/darknet/blob/master/README.md
• ArtyZe's readme: https://github.com/ArtyZe/yolo_segmentation/blob/master/README.md