Comparative Analysis of Single versus Multi-Class Object Detection in Aerial Imagery Using YOLOv8
Abstract
Unmanned Aerial Vehicle (UAV) imagery presents unique challenges for object detection due to extreme variations in object scale, high density, and complex backgrounds. However, existing literature predominantly addresses these challenges by proposing computationally heavy architectural modifications, which are often impractical for the power and edge-computing limitations of UAVs. Furthermore, there is a critical, underexplored research gap regarding how the presence of macro-objects structurally degrades a network’s sensitivity to micro-objects during training. To address this data-centric problem without increasing computational overhead, this study empirically evaluates the performance of the YOLOv8 architecture on the VisDrone2019 dataset, focusing on the trade-offs between single-class and multi-class detection scenarios. We modified the dataset into two configurations: a specialist single-class model targeting only human figures, and a generalist multi-class model targeting both persons and vehicles. Evaluated on the test dataset to measure real-world generalization, the single-class model achieved a mean Average Precision (mAP@50) of 0.520 for the person class. In contrast, the multi-class model demonstrated strong performance on macro-objects like vehicles (mAP@50: 0.893), but experienced a performance drop in the person class (mAP@50: 0.504). This degradation indicates a class interference phenomenon, where the neural network's feature extraction becomes biased toward larger, more distinct objects, causing the False Negative rate for pedestrians to nearly double from 27% to 50%. Despite the accuracy challenges inherent to the VisDrone dataset, both models maintained an exceptional inference speed of 9.0 ms per image on server-grade hardware. These findings demonstrate that YOLOv8 is highly viable for real-time UAV telemetry streaming, but deploying targeted single-class dataset isolation yields superior accuracy for critical micro-object detection tasks.