YOSCA: Confidence Adjustment for Better Object Detection in Aerial Images

Detecting small objects is a significant challenge when evaluating images captured from UAVs. Although YOLO models have been successful in detecting conventional objects, but they still face challenges in detecting small objects in traffic monitoring scenarios. Applying the Slicing-Aided Hyper Inference (SAHI) framework to YOLO models can expand the pixel regions containing small objects, thereby improving the detection of objects with tiny and small sizes. However, this can also lead to the cropping of larger vehicles, which reduces accuracy. This paper proposes a method that combines the results of the YOLO model with an SAHI-enhanced YOLO version, after adjusting the confidence scores of bounding boxes using a Gaussian Mixture Model-based approach, to improve accuracy in detecting small vehicles without compromising the detection of larger objects. This combined approach enhances detection coverage by normalizing the model outputs based on synthesized data, including removing duplicates and score normalization, and then merging the two detection sets. We conduct a thorough evaluation of the proposed method with different versions of YOLO: YOLOv5, YOLOv6, YOLOv7 and YOLOv8. Experimental results on two aerial surveillance datasets, namely, Visdrone and DroneVehicle, demonstrate the increased accuracy of mAP across all tested models.