ACD-DETR: Adaptive Cross-Scale Detection Transformer for Small Object Detection in UAV Imagery

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Small object detection in Unmanned Aerial Vehicle (UAV) imagery presents significant challenges due to complex aerial viewpoints and the difficulty in detecting dense, small targets with indistinct boundaries.
• Existing methods often struggle to preserve fine-grained details and effectively integrate boundary information with semantic features, limiting detection performance.

Purpose of the Study:

• To propose an adaptive end-to-end Transformer detector, ACD-DETR, specifically designed to overcome the challenges of small object detection in UAV imagery.
• To introduce novel modules that enhance feature fusion, boundary calibration, and spatial reasoning for improved detection accuracy and efficiency.
• To present a fusion-enhanced variant, ACD-DETR-SBA+, for scenarios prioritizing detection precision over computational cost.

Main Methods:

• The proposed ACD-DETR framework incorporates three key modules: Multi-Scale Edge-Enhanced Feature Fusion Module (MSEFM) for detail preservation, Omni-Grained Boundary Calibrator (OG-BC) for boundary-aware semantic fusion, and Dynamic Position Bias Attention-based Intra-scale Feature Interaction (DPB-AIFI) for enhanced spatial reasoning.
• A variant, ACD-DETR-SBA+, was developed by removing OG-BC and DPB-AIFI and integrating densely connected Semantic-Boundary Aggregation (SBA) modules to intensify boundary-semantic fusion.
• The models were evaluated on the VisDrone2019 and DOTA datasets.

Main Results:

• ACD-DETR achieved 50.9% mAP@0.5 on the VisDrone2019 dataset, surpassing the RT-DETR-R18 baseline by 3.6 percentage points while reducing parameters by 18.5%.
• The ACD-DETR-SBA+ variant further improved accuracy to 52.0% mAP@0.5, demonstrating the effectiveness of SBA-based fusion.
• Extensive experiments confirmed that ACD-DETR offers a state-of-the-art balance between accuracy and efficiency, with ACD-DETR-SBA+ providing superior performance at a higher computational cost.

Conclusions:

• The proposed ACD-DETR framework effectively addresses the challenges of small object detection in UAV imagery through its novel architectural modules.
• The developed modules, MSEFM, OG-BC, and DPB-AIFI, significantly contribute to preserving details, enhancing boundary awareness, and improving spatial reasoning.
• ACD-DETR and its variant ACD-DETR-SBA+ represent significant advancements in UAV-based object detection, offering flexible solutions for varying accuracy and efficiency requirements.