How to Detect Small Objects using Deep Learning

Abstract

Object detection and classification has become the hall-mark of machine learning and grabbed the attention of computer vision research community for several years. However, detecting small objects in large images still remains an active area of research. Detecting small objects is challenging because of small subject area on image and few unique features of the object. Moreover, due to limited annotated data availability, it is difficult to employ machine learning algorithms. This is specially true for remote sensing applications. Detecting cars in satellite imagery can help in traffic monitoring, urban planning and surveillance. To tackle these challenges, we employ data augmentation using image blending, and style blending using Generative Adversial Networks (GAN) to augment subject top-view car images in satellite data and create large datasets which otherwise may require substantial human effort. The augmented dataset is utilized to train standard object detection Convolutional Neural Networks (CNN) and benchmark their performance on standard datasets.