Algorithm evaluation demonstrates enhanced detection accuracy in aerial imagery, indicating improved handling of complex backgrounds and variable angles.
Key Points
The objective was to improve oriented object detection in remote sensing imagery by mitigating interference from cluttered backgrounds and handling arbitrary object rotation angles.
Engineered HRDet, a one-stage oriented object detector integrating a hierarchical mask prompting module and a robust integrated regression strategy.
Developed a semantic mask prediction branch paired with hierarchical Softmax to separate objects from background clutter and route differentiated features across adjacent layers.
Formulated an oriented IoU loss measuring discrepancies across central point distance, side length, and angle, evaluated on the DOTA-v1.0, DOTA-v2.0, DIOR-R, and HRSC2016 aerial datasets.
HRDet significantly improved detection accuracy compared to refine-stage counterparts across four standard aerial benchmarks.
The detector maintained the computational speed advantages characteristic of single-stage architectures while overcoming angle-invariance limitations in conventional IoU losses.