ASCDet: cross-space UAV object detection method guided by adaptive sparse convolution

Abstract

UAV object detection, a critical aspect of remote sensing applications, faces challenges due to high object sparsity and complex backgrounds, leading to excessive computational demands. To address these issues, we propose the Cross-Space UAV Object Detection Method Guided by Adaptive Sparse Convolution (ASCDet), a more efficient solution for UAV remote sensing object detection. ASCDet introduces a plug-and-play detection head compatible with various detection frameworks, significantly reducing computational costs. The method utilizes an adaptive pixel-level mask unit based on a task-alignment strategy to accurately localize object regions. These masks guide cross-space object detection through sparse convolutions, while a global context enhancement strategy within the sparse convolution module enriches the contextual information, maintaining detection accuracy. Extensive experiments on the VisDrone and UAVDT datasets, comparing ASCDet with benchmark methods such as Faster R-CNN, RetinaNet, FSAF, GFL V1, and TOOD, show that ASCDet improves AP[Formula: see text] by up to 5.8%, increases FPS by up to 34.3%, and reduces GFLOPs by approximately 80%. Additionally, ASCDet enhances detection performance for anchor-free methods, demonstrating superior accuracy and computational efficiency. These results highlight ASCDet's effectiveness in improving detection accuracy, computational efficiency, and sparse detection performance in UAV remote sensing tasks.