Title: Object Detection in Aerial Images: What Improves the Accuracy?

Abstract: Object detection is a challenging and popular computer vision problem. The problem is even more challenging in aerial images due to significant variation in scale and viewpoint in a diverse set of object categories. Recently, deep learning-based object detection approaches have been actively explored for the problem of object detection in aerial images. In this work, we investigate the impact of Faster R-CNN for aerial object detection and explore numerous strategies to improve its performance for aerial images. We conduct extensive experiments on the challenging iSAID dataset. The resulting adapted Faster R-CNN obtains a significant mAP gain of 4.96% over its vanilla baseline counterpart on the iSAID validation set, demonstrating the impact of different strategies investigated in this work.