Visión Artificial y Grafismos: Project 3 part 2 Mask / No mask, train a Yolov4 cnn with Darknet

Project 3 part 2"Mask / No mask", trained with Darknet and Yolov4.cfg

Task 2: Train the object detector using Yolo v4 [30 Marks]

You should repeat the above experiment using the Yolo v4 architecture and model.

HINT:

1. You need to change Step 6 along with the above changes. You can use the following link as a reference for training the model using Yolo v4 architecture and pre-trained model.

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

I wrote a notebook with the basic pipeline for train a yolov4 cnn with Darknet in Google colab, this are the steps:

1.- I created a folder in Google drive to upload the test files.

2.- Mount a google drive and check its files.

3.- Clone, configure and compile darknet.

4.- Configure yolov3.cfg file.

5.- Create names and data files and save to Google drive.

6.- Create a folder where unzip the image dataset.

7.- Create a train.txt file.

8.- Download pre-trained weights for the convolutional layers file.

9.- Start training.

10.- Object detection with the test images. ...OOPS!

I got the weights after 12 hours training with the best "activation: mist" in .cfg file

Unsuccesfully I tried to install a version of OpenCV newer than 4.1.2 in Google Colab, I didn't get the trick.

I know that OpenCV 4.3.0 supports "activation: swish" that is almost like "activation: mish", I also tried that here after I modified "yolov4_testingb.cfg".

OpenCV 4.4.0 supports "activation: mish" in function 'ReadDarknetFromCfgStream', Then I run the tests yolov4 with OpenCV 4.4.0 in my own system, here are the outputs.

I found that yolov4 takes longer to train, but is faster and more precise than yolov3.

10.- Object detection with the test images.