ROBUST BOUNDING BOX REGRESSION FOR SMALL OBJECT DETECTION

Deep learning advances have propelled common object detection development. However, small object detection in aerial images remains inaccurate. Intersection-over-Union (IoU) has limitations in handling the separation or inclusion cases of prediction and ground truth boxes, especially in small object detection, which significantly hampers the process of bounding box regression. To tackle the issue, we propose Balanced Corner-IoU (BC-IoU) loss, which incorporates both corner point distances and IoU metric, while simultaneously introducing the instance area as a component of loss terms. Moreover, Point Offset Module (POM) branch is developed to generate additional positive samples for small object regression by dynamically controlling anchor point generation. With the above designs, Scale Adaptive Network (SAN) provides a solution to bounding box regression of small objects. Experiments on the small object detection dataset show that BC-IoU loss outperforms other IoU loss variants and that SAN significantly improves performance over the Fully Convolutional One-Stage (FCOS) baseline.