Object Detection in Aerial Images with Uncertainty-Aware Graph Network

In this work, we propose a novel uncertainty-aware object detection framework with structured-graph, where nodes and edges are denoted by objects their spatial-semantic similarities, respectively. Specifically, aim to consider relationships among for effectively contextualizing them. To achieve this, first detect then measure semantic spatial distances construct an graph, which is represented graph neural network (GNN) refining visual CNN features objects. However, results of every inefficient may not be necessary, as that include correct predictions low uncertainties. Therefore, handle uncertain only transferring the representation from certain (sources) (targets) over directed but also improving on regarded representational outputs GNN. Furthermore, calculate training loss giving larger weights objects, concentrate while maintaining high performances We refer our model Uncertainty-Aware Graph DETection (UAGDet). experimentally validate ours challenging large-scale aerial image dataset, namely DOTA, consists lots small large sizes in image, improves performance existing network.