Scholarlink

This paper presents a thorough evaluation of three representative object detection models—YOLOv5s, SSD300, and Faster R-CNN—focusing on performance, efficiency, and benchmarking metrics. We assess each model’s accuracy (using mAP, precision, and recall), inference speed (FPS and derived boxes per second), and computational efficiency (including proposed metrics BPS and EAS). Experiments on standard datasets (COCO and VOC) demonstrate clear trade-offs: YOLOv5s achieves real-time processing (90 FPS, 37% COCO mAP), SSD300 balances speed and lower accuracy (58 FPS, 24% COCO mAP), and Faster R-CNN attains high accuracy ( 40% COCO mAP) with lower throughput (7 FPS). We apply composite metrics—Boxes per Second (BPS) and Energy-Accuracy Score (EAS)—to evaluate speed-accuracy-energy trade-offs. Real-world deployment implications are discussed, with practical insights for edge devices such as Raspberry Pi and Jetson Nano. This benchmarking framework supports application-specific detector selection in domains like surveillance, UAVs, and mobile vision

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking the Trade-Offs in Object Detection: Accuracy, Speed, and Energy Efficiency

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Abhnish Kumar,
  • Manan Wadhwa,
  • Dinesh Kalla,
  • Santhosh Chandra Konduru,
  • Chinmay Nandawat,
  • Moolchand Sharma

摘要

This paper presents a thorough evaluation of three representative object detection models—YOLOv5s, SSD300, and Faster R-CNN—focusing on performance, efficiency, and benchmarking metrics. We assess each model’s accuracy (using mAP, precision, and recall), inference speed (FPS and derived boxes per second), and computational efficiency (including proposed metrics BPS and EAS). Experiments on standard datasets (COCO and VOC) demonstrate clear trade-offs: YOLOv5s achieves real-time processing (90 FPS, 37% COCO mAP), SSD300 balances speed and lower accuracy (58 FPS, 24% COCO mAP), and Faster R-CNN attains high accuracy ( 40% COCO mAP) with lower throughput (7 FPS). We apply composite metrics—Boxes per Second (BPS) and Energy-Accuracy Score (EAS)—to evaluate speed-accuracy-energy trade-offs. Real-world deployment implications are discussed, with practical insights for edge devices such as Raspberry Pi and Jetson Nano. This benchmarking framework supports application-specific detector selection in domains like surveillance, UAVs, and mobile vision