Region Proposal Networks (RPN) Enhanced Slicing for Improved Multi-Scale Object Detection

Detecting objects in high-resolution images, particularly those from advanced drones and satellites, presents significant challenges to traditional object detection methods. These challenges are amplified when dealing with multi-scale objects, ranging from very small to larger sizes. Traditional slicing-aided methods like SAHI have shown proficiency in enhancing the detection of smaller objects but often struggle with larger object detection due to the limitations of fixed-step sliding windows. In this work, we introduce the Enhanced Slicing-Aided Inference Framework (ESAHI), a novel approach that integrates Region Proposal Networks (RPN) into the slicing process. This integration effectively overcomes the constraints of fixed-step slicing, thereby improving the detection accuracy across multiple object scales. ESAHI utilizes RPN to generate object-centric regions in high-resolution images, creating additional patches that retain complete features of larger objects. These patches are processed alongside those obtained through traditional sliding windows, ensuring comprehensive coverage of the imagery. Our comprehensive evaluations on the VisDrone2019 dataset demonstrate that ESAHI significantly enhances object detection performance. Specifically, ESAHI increases object detection average precision (AP50) for medium-sized objects by 1.5%, 1.6%, and 4.7%, and for larger objects by 7.3%, 8.2%, and 22.3% across different slicing sizes of 640*640, 480*480, and 320*320, respectively.