Object Detection in High Resolution Aerial Imagery Using Detection Transformer

Object detection in high-resolution aerial imagery has received much attention nowadays due to its applications in geosciences, urban planning, disaster management, and surveil- lance. However, there exist challenges such as scale variation, cluttered backgrounds, occlusions, and less annotated datasets. Traditional CNNs have shown great promise, yet they fail to detect long-distance dependencies and complicated spatial relationships. This paper evaluates the function of DETR for object detection in aerial images. Unlike CNN-based detectors that depend on region proposal networks and anchor-based methods, DETR depends on a full end-to-end transformer architecture along with a direct set prediction method that removes the requirement for hand-designed priors. With extensive experiments carried out on datasets like Airbus Aircraft, Rare Planes, and DOTA, observations show that DETR performs better with mAP scores that are as much as 18% higher than ResNet-based architectures. Fur- Furthermore, we propose a hybrid model that is DETR-CNN, which partners both the strength of feature extraction from CNNs and the global attention mechanisms in DETR, thereby improving the accuracy of detection on both Horizontal and Oriented Bounding Box detections. Our results show that transformer-based models are most effective in aerial object detection, which bodes well for remote sensing, autonomous surveillance, and disaster response applications. This study presents an end-to-end DETR-based method for object detection in aerial imagery, demonstrating improvements in accuracy and simplicity over traditional methods.