This repository contains ball detection classifier using darknet YOLO v3 model for URC 2019 competition.
No. of Classes: 1
Name of the Class: Ball
You need to install darknet on your Ubuntu 18.04 machine. Follow the tutorial here to install darknet in order to use YOLO V3/v4. Check the installation with your webcam to reassure that you have installed it properly.
Training & Validation Dataset (i.e train_ball & valid_ball folder) with proper labeling should be kept inside ~/darknet/data directory. You can download the training and validation dataset from here. Download pre-trained weight file for convolutional layer from darknet53.conv.74 & keep it inside ~/darknet directory.
In order to label data collected from different sources, install labelImg. You can find the corresponding download & installation proccedure here. Just remember when using 'labelImg', change the image labeling format from pascal to YOLO (by clicking on the side button 'pascal') and save the corresponding 'img.txt' labeled files accordingly. Each labeled image will produce a 'x.txt' file where the value 'x' will be the name of the original image.
To create train_ball.txt and test_ball.txt, download train_txt_creator.py & test_txt_creator.py files from this repository, put them inside Home directory and run in the terminal:
## To create train_ball.txt file, don't forget to change the path mentioned inside train_txt_creator.py file
python train_txt_creator.py
## To create test_ball.txt file, don't forget to change the path mentioned inside test_txt_creator.py file
python test_txt_creator.py
cd ~/darknet
./build-release/darknet detector train data/train_ball.data cfg/yolov3_train_ball.cfg darknet53.conv.74
The default weight file will be saved after every 1000 iterations inside ~/darknet/backup directory during training. You can use your own weight file or you can use mine. In my case, I have trained for around 2000 iterations. It took around 6 hours for training purpose. You can download the weight file from here.
cd ~/darknet
./build-release/darknet detector demo data/train_ball.data cfg/yolov3_test_ball.cfg backup/yolov3_train_ball_2000.weights -c 0
The data were trained inside Ubuntu 18.04 with Nvidia Graphics GTX 1070, Driver version 440.82 & CUDA version 10.2. I took around 300 images for one class of which 270 (90%) images were used as training dataset & 30 images (10%) were used as validation dataset. The testing was done with a logitech webcam and the detection frame rate was 15 fps.
96% images used in this work during training are taken from mobile phone & other 4% used are taken from internet. Respective people can claim copyright for those images.