Discriminative features enhancement for low-altitude UAV object detection

Object detection is a pivotal task in low-altitude UAV application. Here the small scale objects are dominant due to shooting distance and angle and insufficient feature information due to the data from real world scenes. Although general detector has made great progress, it is not suitable for small scale object detection directly. Dense detector has potential because of the pixel-by-pixel detection but the resolving power of complex background and objects especially small scale objects is still insufficient. We propose a Feature Guided Enhancement module by designing two non-linear learning operators to guide more discriminative features when training. Further, a Scale-Aware Weighted loss function is proposed to dynamically weight the loss of various scale objects by statistical computing and highlight the contribution of small scale objects. Experimental results show that our method can effectively improve FCOS and ATSS, and our models obtain better performance by 1.5% and 0.6% AP respectively on VisDrone 2018 dataset.