Tutorial YOLO v4 to detect wheat heads from crops. The project is based on this Kaggle Competition: https://www.kaggle.com/c/global-wheat-detection.
In this competition, you’ll detect wheat heads from outdoor images of wheat plants, including wheat datasets from around the globe. Using worldwide data, you will focus on a generalized solution to estimate the number and size of wheat heads. To better gauge the performance for unseen genotypes, environments, and observational conditions, the training dataset covers multiple regions. You will use more than 3,000 images from Europe (France, UK, Switzerland) and North America (Canada). The test data includes about 1,000 images from Australia, Japan, and China.
These examples have wheat heads detected
These examples don't have any wheat heads
The data given in the contest have some problems:
1. The given label format is not YOLO v4 format:
The given is [x_min y_min width height], while YOLO v4's is [class x_center y_center width height].
So we have to convert into YOLO v4 format:
# Convert bounding box to YOLO v4
def convert_to_YOLOv4(list_of_annotation):
x_min = list_of_annotation[0]
y_min = list_of_annotation[1]
width = list_of_annotation[2]
height = list_of_annotation[3]
width_img = 1024
height_img = 1024
# Conversion
bx = round((x_min+width)/(2*width_img),6)
by = round((y_min+height)/(2*height_img),6)
w = round(width/width_img,6)
h = round(height/height_img,6)
annotation = [0,bx,by,w,h]
return annotation
2. Not all training data have labels:
The number of images in "train.csv" file is less than the in "train" folder
So we can only utilize images in "train.csv" and regretly waste the remnants
3. There are only training and test data, missing validation data:
So we have to split training and validation data in ratio 80/20
# Training/test data split
from sklearn.model_selection import train_test_split
image_id_train, image_id_test, bbox_train, bbox_test = train_test_split(image_id, bbox, test_size=0.2, random_state=42)
Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.
Source code: https://github.com/AlexeyAB/darknet
The result of YOLO v4 is impressive compared to other models:
