A Novel Keypoint Supplemented R-CNN for UAV Object Detection
Justin Butler, Henry Leung
Abstract
Aerial imagery and remote sensing applications are excellent examples of environments that challenge current deep learning-based object detection architectures due to the images generally consisting of small objects within large, complex backgrounds. In this article, the region proposal network (RPN) of the successful Mask region-based convolutional neural network (R-CNN) architecture is supplemented with a second nonconvolutional branch to increase the number of accurately predicted regions of interest (RoIs). The additional RoIs are predicted using keypoint features unique from the Mask R-CNN backbone which better capture the existence of small objects and allow for a set of RoIs to be predicted independent of the standard Mask R-CNN feature pyramid network (FPN). The proposed architecture is evaluated on synthetic data to demonstrate the improved performance of the supplemented RPN on smaller objects and objects of changing scale. The two-branch network is then demonstrated to achieve improved detection accuracy when applied to the challenging aerial-cars and vehicle-detection datasets (VDDs).