Finding Instances of Objects in Aerial Images using Earth Vision Small Scale Dataset

Object detection has made enormous progress in naturalistic images during the last decade. However little has been done to detect miniscule objects in aerial images. Because of their small size and complex backstory detecting these small objects becomes very difficult. RCNN, Fast RCNN, Faster RCNN, YOLO, and SSD are some of the object detection algorithms utilized. Many changes have been made to these algorithms in order to enhance performance both in speed and accuracy. YOLOR i.e. You Only Learn One Representation is one such algorithm that has been developed specifically for object detection, unlike other algorithms which include object classification or object analysis. YOLOR is a one-stage detector, capable of performing numerous tasks at a time. In this paper, we present a brief summary of various object detection algorithms. We also present our dataset ‘Earth Vision small-scale dataset' to make tiny object detection easier. The dataset consists of 700 highresolution images. We also studied the performance of the YOLOR algorithm on this dataset. Experimental results show that the accuracy obtained is 76% which is good due to the small size of the dataset and the tiny nature of the objects. The source code of this work is at: https://github.com/PawarNehaAshok/Yolor-EarthVision-Small-Scale-Dataset.