An Empirical Study of a Shape-Adaptive Approach in Aerial Object Detection

Aerial object detection in real-world scenarios is always an important and challenging task, with numerous applications including resource monitoring, environmental assessment, military support, and urban management. The goal of this task is to identify and locate objects in various types of images, including satellite images, radar data, and photos captured by drones. Label assignment is a crucial and widely used strategy in object detection tasks. However, it presents several challenges, such as the imbalance between positive and negative samples and the sensitivity to IoU (Intersection over Union) thresholds. We researched a novel dynamic shape-adaptive selection strategy, which selects positive samples according to the shape and characteristic distribution of objects. This strategy has proven effective in overcoming the challenge of positive sample selection by utilizing dynamic IoU thresholds. We conducted experiments using the Shape-Adaptive Selection and Measurement (SASM) method, an approach based on oriented bounding boxes, to detect objects. The three datasets used in this research are DOTA-Ship, FGSD2021, and Drone Vehicle. The experimental results of SASM on the Drone Vehicle dataset demonstrated noteworthy performance with a mAP of 61 %, outperforming Faster R-CNN by 16.41 % and RoITransformer by 13.09%.