RFNet: Toward High-Quality Object Detection in Aerial Images

Due to the object overlap and multi-scale transformation problems in aerial image detection, the detection system has false and missed detections. In this paper, we propose feature Re-Fusion Network, named RFNet. Specifically, the recurses multi-layer feature maps into a pyramid network, cross fusion of upper and lower features to enhancing information flow. Next, we introduced a lightweight module that connects feature semantic information for fusion to improve the feature fusion ability. Finally, we optimized the parameters of attention mechanism block to adapt to aerial images detection. We equipped it to the modern detector Cascade R-CNN on the VisDrone dataset, which improved the mean Average Precision(mAP) by 1.1%, increased middle-size object detection accuracy by 2.1%, and increased the accuracy of the truck class by 2.4%.