Abstract
Object detection in aerial images is continuously studied for various purposes such as national security, disaster monitoring, and meteorological observation. It is difficult to improve recent object detection methods based on a single model using deep learning due to severe class imbalance. This paper proposes a deep ensemble method combining two models with different strengths and a class-dependent thresholding method by considering the object distribution. We demonstrate the superiority of our methods in a series of experiments. In addition, we take 1st place in both public and private scores in the Arirang satellite image AI object detection contest.
Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
