A Remote Sensing Image Object Detection Algorithm Based on a Fusion of Classification and Regression Branches

The detection head of most object detection algorithms obtains the category information and location information through classification branch and regression branch, respectively. Usually, the two branches do not influence each other, therefore neither the positional information implied in the classification branch, nor the classification information implied in the regression branch is used. In this paper, we redesign the traditional detection head to solve the above problems. The mAP of our method in the public remote sensing image object detection dataset DOTA-v1.0 reaches 77.27%, while the mAP of baseline (RoI Transformer) is only 75.43% under the same training conditions. The test results suggest that the detection head branch fusion algorithm can improve the performance of the remote sensing image object detection model.