Comparative Evaluation of Anchor-Based YOLOv11 and Anchor-Free FCOS for Small-Object UAV Detection

Authors : Hilal İkra Yücel, İlyas Özer, Adem Dalcalı

Pages : 214-221

Doi:10.46387/bjesr.1739026

View : 172 | Download : 360

Publication Date : 2025-10-27

Article Type : Research Paper

Abstract :In this study, low-cost image-based detection of Unmanned Aerial Vehicle (UAVs) over critical infrastructure and urban areas is investigated. Traditional hand-crafted feature methods (Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT), Local Binary Patterns (LBP)) and single-stage classifiers (Support Vector Machines (SVM), Adaptive Boosting (AdaBoost)) are shown to struggle with small targets under dynamic conditions. To address this, anchor-based YOLOv11 and anchor-free Fully Convolutional One-Stage (FCOS) architectures are compared on the “UAV Drone” Kaggle dataset. Both models use a unified preprocessing pipeline (resize to 640×640, normalization, data augmentation) and three-fold cross-validation for 50 epochs with the AdamW optimizer and a OneCycleLR schedule. Results reveal that YOLOv11 achieves 66.6 % mAP@[0.5–0.95] at ~10 FPS, while FCOS attains 64.1 % mAP at ~20 FPS with lower memory use. Thus, YOLOv11 is recommended for high-accuracy research, and FCOS variants for real-time, resource-constrained applications.
Keywords : İHA Tespiti, Ankraj Tabanlı Nesne Tespiti, Tam Konvolüsyonel Tek Aşamalı

ORIGINAL ARTICLE URL