Indonesia Comparative performance analysis of YOLOv8 small and larger for real-time website-based monitoring

Faiz Noor Adhytia Adhytia, Raka Surya Kusuma, Vincentius Raditya Agung Soedomo, Yonathan Purba SantosaNon-Scopus2 Februari 2026Afiliasi: ID

Abstrak

River trash pollution in Indonesia demands efficient real-time monitoring solutions. By using deep learning, while YOLOv8 is a promising model for implementation, it is a family of models with each variant differ in accuracy, speed, and computational demand. Among these, YOLOv8s and YOLOv8l come out as potential variants due to the balance between speed, accuracy, and computational demand. Therefore, to address this gap, this study intends to compare the YOLOv8 small (YOLOv8s) and YOLOv8 large (YOLOv8l) variants for real-time, website-based river trash monitoring systems, aiming to identify the optimal balance between accuracy and inference speed for practical real-time deployment. A combined dataset of 66 images, consisting of 86% images from Kaggle’s 2024 dataset and 14% AI-Generated images from Krea AI, was augmented using Roboflow to produce 591 annotated images. Both models were treated equally with the same dataset and method. The evaluation was conducted using Precision, Recall, mAP50, mAP50-95, Inference Speed, and Frame per Second (FPS) speed. Consequently, YOLOv8s exceeded YOLOv8l, achieving approximately 20% higher precision, 30% better detection quality, and 20% higher mAP across IoU threshold, and 150%-340% faster FPS. This finding indicates that YOLOv8s offer better accuracy and speed trade-off for real-time implementation than YOLOv8l. Moreover, successful integration to the website with real-time stream and threshold alert, confirm its feasibility for proactive waste and flood management.

Identifier

Ekspor BibTeX
Scroll to Top