Fine-grain uncommon object detection from satellite images

Abstract

The ever increasing amount of earth observing satellite images is a vast treasure trove of interesting objects. We address the topic of object detection from satellite images in cases where the object is rarely observed and hence there is very little availability of images to support training classifiers. Unlike objects observed on the ground, there is no equivalent ImageNet with labeled data for objects as seen from satellite or aerial platform sensors that could be used to train classifiers. In addition, we focus on specific uncommon objects with very limited observations. To overcome the lack of training data, we built a near-class object detector and verified the uncommon object detection using images from different domains. We demonstrate the performance of our uncommon object detector and show a high detection rate in satellite images.