Enhanced lightweight detection of small and tiny objects in high-resolution images using object tracking-based region of interest proposal

Abstract

Detecting small objects in high-resolution images is challenging in practical applications. Downsampling the image makes small objects difficult or impossible to detect, while processing multiple low-resolution detection windows using a sliding-window approach is both time-consuming and impractical. To deal with this engineering problem, we present a novel window-based object detection system designed for detecting tiny and multi-scale objects in high-resolution images. Our approach introduces two Region of Interest (ROI) Modules that select full-resolution areas for the detector to focus on. We use a low-resolution current frame to estimate ROIs with a segmentation-based branch combined with past detection metadata to predict object locations with a tracking-based branch. By fusing outputs from these modules, we effectively recover regions with tiny objects overlooked by the estimation branch, and we significantly reduce the number of object detection runs compared to the sliding-window approach, maintaining the method’s speed. Moreover, we employ a fusion of downscaling and sliding-window techniques within large ROIs, complemented by our novel Overlapping Box Suppression (OBS) algorithm to reduce partial false-positive detections. We analyze our system and all its components on two challenging datasets — SeaDronesSee and DroneCrowd to show superior performance compared to state-of-the-art object detectors. Our approach enhances both the Artificial Intelligence (AI) and engineering domains by improving the quality and efficiency of tiny object detection, facilitating its integration into demanding real-time robotics applications. The inference code is available at https://github.com/deepdrivepl/TinyROIFusion.