Small Object Difficulty (SOD) Modeling for Objects Detection in Satellite Images

Abstract

Recent increases in aerial image access and volume, increases in computational power, and interest in applications have opened the door to scaling up object detection to production. Aerial data sets are very large in size, and each frame of the data set contains a huge number of dense and small objects. Deep learning applications for aerial imagery are behind due to a high variety between datasets (e.g. object sizes, class distributions, object feature uniformity, image acquisition, distance, weather conditions), the size of objects in satellite imagery, and the subsequent failure of state-of-the-art architecture to capture small objects, local features, and region proposals for densely overlapped objects in satellite images. In this paper, we provide a novel pipeline that improves the back-end through spatial pyramid pooling, a partial cross-stage network, a region proposal network via heatmap-based region proposals, and object localization and identification through a novel image difficulty score t hat adapts the Overall focal loss measure based on the image difficulty. Our proposed model outperformed the state-of-the-art method in mAP by 1.8% and 2.3% in the DOTA and DIOR data sets, respectively.