Object Detection Dataset for Overhead Images of Wind Turbines

Overview This project contains three object detection datasets for overhead imagery of wind turbines. Overhead Imagery of Wind Turbines contains real images and their labels. The other two of these datasets contain synthetic images and their labels. Contents Each of these datasets contains a set of images and a label corresponding to each image. The images are 608x608 pixels. The labels have the same names as their corresponding image (e.g. image_123.jpg corresponds with the label image_123.txt). Each label contains information about the ground truth bounding boxes for the wind turbines in each image. Each line in the labels contains information about the bounding box for a wind turbine, formatted as class x y height width (YOLOv3 format), where x and y are the coordinates for the center of the bounding box. These values (except for class value) are relative proportions to the size of the image, varying from 0-1. For more details on these datasets (their overview, use, reason for being, and methods), read the description for the specific datasets. Experimentation The primary experiment done with this data was training a YOLOv3 model on the real images contained in Overhead Imagery of Wind Turbines and then testing on 20% of this data to get baseline results. Then, the synthetic data was added to the training set to train another model, and the performance was evaluated on the same testing set of real overhead images. If you would like to check out the experimentation, go to https://dataplus-2020.github.io/. We also have a github repo at https://github.com/dataplus-2020/yolov3_wnd_code, but be warned the yolov3 repo is very messy. The work is still being done as a Duke Bass Connections project, which can be seen at the figshare here: https://figshare.com/projects/Adding_Synthetic_Imagery_for_Object_Detection_on_Overhead_Images_of_Wind_Turbines/96131 and our github organization here: https://github.com/Duke-BC-DL-for-Energy-Infrastructure.