GA-Net: Accurate and Efficient Object Detection on UAV Images Based on Grid Activations

Abstract

Object detection plays a crucial role in unmanned aerial vehicle (UAV) missions, where captured objects are often small and require high-resolution processing. However, this requirement is always in conflict with limited computing resources, vast fields of view, and low latency requirements. To tackle these issues, we propose GA-Net, a novel approach tailored for UAV images. The key innovation includes the Grid Activation Module (GAM), which efficiently calculates grid activations, the probability of foreground presence at grid scale. With grid activations, the GAM helps filter out patches without objects, minimize redundant computations, and improve inference speeds. Additionally, the Grid-based Dynamic Sample Selection (GDSS) focuses the model on discriminating positive samples and hard negatives, addressing background bias during training. Further enhancements involve GhostFPN, which refines Feature Pyramid Network (FPN) using Ghost module and depth-wise separable convolution. This not only expands the receptive field for improved accuracy, but also reduces computational complexity. We conducted comprehensive evaluations on DGTA-Cattle-v2, a synthetic dataset with added background images, and three public datasets (VisDrone, SeaDronesSee, DOTA) from diverse domains. The results prove the effectiveness and practical applicability of GA-Net. Despite the common accuracy and speed trade-off challenge, our GA-Net successfully achieves a mutually beneficial scenario through the strategic use of grid activations.

Keywords

• drone-view object detection
• real-time inference
• background bias mitigation