Mitigating the Effect of Resizing in YOLO for Small Object Detection Applied to Raptor Detection

Enol GARCÍA GONZÁLEZ, Laura MENÉNDEZ GARCÍA, José R. VILLAR, Javier SEDANO

Abstract

Single-stage object detectors such as YOLO are widely used due to their efficiency and simplicity, but they feature well-known limitations when dealing with small object detection in high-resolution images. A key factor contributing to this issue is the global resizing of input images, which can severely reduce the effective spatial resolution of small targets. This problem is especially relevant in outdoor monitoring scenarios, such as the detection of birds of prey in free-range poultry farms, where the objects of interest often appear distant and occupy only a small fraction of the analysed scene. This paper proposes a region-based preprocessing strategy which was designed in order to mitigate the adverse effects of image resizing in the context of YOLO-based detectors. This approach consists in partitioning the original image into multiple localized regions that are processed independently by the detection model, allowing small objects to be analysed at a higher relative resolution without altering the network architecture or the training procedure. This method is intended to be model-agnostic and easily integrable into the existing YOLO-based pipelines. The proposed strategy is applied for detecting birds of prey in outdoor farm environments, providing a structured framework for improving small object representation in single-stage detectors. This work aims to contribute a practical preprocessing solution that addresses a fundamental limitation of YOLO in scenarios dominated by small and sparsely distributed objects.

Keywords

Object detection, Bird detection, YOLO.

View full article