CNN Adaptations for Boat Detection in Aerial Images Tested on Yolo v2

In this article, we focus on boat real-time detection in aerial images taken by UAVs Unmanned Aerial Vehicles. Several methods dealing with this problem are based on convolutional networks. Generally, they start from existing networks that have demonstrated their effectiveness with datasets like COCO Common Object in Context [1] and adapt them to improve their performance on aerial images. The adaptations made should not cause a higher execution time than the initial network, especially if the image processing and detection must be done in real-time. They must also participate in the increase in Recall by detecting even small objects on aerial images which is the case of most objects present in photos taken at high altitude. The purpose of this article is to test the effectiveness of certain adaptations with our boat dataset. We will also propose new adaptations. The tests will be performed using the Yolo v2 neural network [2].