SF-DETR: A Scale-Frequency Detection Transformer for Drone-View Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Drone-based object detection faces challenges with tiny objects, complex backgrounds, and scale variations.
• Existing methods often lack computational efficiency when addressing these drone-view complexities.

Purpose of the Study:

• To develop a novel end-to-end framework, Scale-Frequency Detection Transformer (SF-DETR), for improved drone-based object detection.
• To address challenges such as tiny objects, scale variations, and feature detail loss in aerial imagery.

Main Methods:

• Proposed SF-DETR framework featuring a lightweight ScaleFormerNet backbone.
• Incorporated Dual Scale Vision Transformer modules and a Bilateral Interactive Feature Enhancement Module.
• Utilized a Multi-Scale Frequency-Fused Feature Enhancement Network for robust feature representation.

Main Results:

• SF-DETR achieved 51.0% mAP50 and 31.8% mAP50:95 on the VisDrone2019 dataset.
• Outperformed state-of-the-art YOLOv9m by 6.2% and RTDETR-r18 by 4.0% in key metrics.
• Demonstrated strong generalization capabilities on the HIT-UAV dataset.

Conclusions:

• SF-DETR establishes a new benchmark for drone-view object detection performance.
• The lightweight architecture is suitable for real-world aerial surveillance on embedded devices.