Multi-scale Aerial Object Detection Using Feature Pyramid Networks

Abstract

Aerial object detection on an UAV or embedded vision platform requires accurate detection of objects with various spatial scales and has numerous applications in surveillance, traffic monitoring, search, and rescue, etc. The task of small-object detection becomes harder while using standard convolutional neural network architectures due to the reduction in spatial resolution. This work evaluates the effectiveness of using feature pyramid hierarchies with the Faster R-CNN algorithm for aerial object detection. The VisDrone aerial object detection dataset with ten object classes has been utilized to develop a Faster R-CNN ResNet model with C4 and FPN architectures to compare the performance. Significant improvement in the performance obtained by using feature pyramid networks for all object categories highlights their importance in the multi-scale aerial object detection task.