| Autors: Ismailov, A. V., Hristov, V. D. Title: Pruning YOLOv8 Detection Models Using Sipp Pruning Method Keywords: Computational Efficiency, Edge AI, Model Pruning, Object Detection, Sensitivity-Informed Pruning (SiPP), YOLOv8 Abstract: Recent advancements in deep learning have led to increasingly complex neural networks, posing challenges for deployment on resource-constrained devices. This paper investigates the Sensitivity-Informed Provable Pruning (SiPP) algorithm to compress YOLOv8n and YOLOv8s object detection models while maintaining competitive performance. We evaluate SiPP's efficacy across four datasets-COCO, KITTI, Car Object Detection, and PlantDoc-pruning models at 20%, 40%, 50%, 60%, and 80% keep ratios (KR). Results demonstrate that 20% pruning (80% KR) retains near-original accuracy (e.g., mAP50-95, drops from 44.9 to 40.6 for YOLOv8s on COCO), with steeper declines beyond 60% KR. Notably, models pruned on domain-specific datasets (e.g., KITTI) show greater robustness to aggressive pruning. Our findings highlight SiPP's potential to enable efficient deployment of YOLOv8 in edge applications like autonomous driving and precision agriculture, with minimal performance trade-offs. References
Issue
|
Вид: публикация в международен форум, публикация в реферирано издание, индексирана в Scopus