A New GNN-Based Object Detection Method for Multiple Small Objects in Aerial Images

Detecting small objects in aerial images is challenging due to limited information of the object and complex backgrounds surrounding them. Most of the existing object detection methods focus on detecting individual objects and have different feature extraction methods based on object appearance. In this paper, we propose a new Graph Neural Networks (GNN) based method to refine detection results generated by object detectors. In this method, we construct a detection graph by using the predicted detection bounding boxes as nodes, while the features of a bounding box become its node features. Edges are added using distance and topological information. Then, based on the detection confidences of the bounding boxes, some nodes are labeled as bird or non-bird, where nodes with high detection confidences are labelled bird and nodes with very low detection confidences are labeled as non-bird. GNN algorithms are trained using the labeled portion of the graph to infer the classes of the unlabeled nodes. Our experimental results on detecting waterfowl in aerial images show that the new method significantly improved detection accuracy and robustness by significantly reducing false positives.