ContextTiny-Net: An Ultra-Tiny Object Detection Network for UAV Aerial Images in Urban Scenarios

In the intelligent transportation system of smart cities, object detection from UAV aerial imagery serves as the core technical support for traffic flow monitoring, violation detection, and emergency response. However, traffic objects captured from UAV perspectives typically exhibit extremely low pixel occupancy and are embedded in complex backgrounds, leading to three fundamental limitations in existing detection methods: insufficient utilization of global context information, inaccurate weak feature enhancement, and severe feature scale confusion. To address these challenges, this paper proposes ContextTiny-Net, an ultra-tiny object detection network built upon multi-dimensional symmetry design principles for urban UAV scenarios. Specifically, we first construct a global–local perception symmetric MetaFormer backbone and a hierarchical scale symmetric four-layer detection head, which achieves full-coverage detection from ultra-tiny to regular traffic objects with minimal computational overhead. Second, we design an information-theoretic and spatial-distribution-complementary symmetric-weak feature enhancement module, which accurately locates and strengthens weakly activated regions of small objects from two mutually complementary and symmetric dimensions. Finally, we propose a cross-scale decoupling symmetric feature fusion module and a symmetric Gaussian distribution-based normalized Wasserstein distance loss, which effectively eliminate scale confusion and significantly improve the robustness of small object bounding box regression. Extensive experiments on three mainstream benchmarks (AI-TOD, VisDrone, and COCO) demonstrate that ContextTiny-Net outperforms state-of-the-art methods in both overall detection accuracy and ultra-tiny object detection performance, verifying the effectiveness of the proposed symmetry-enhanced design paradigm.