Comparative Analysis of Machine Learning Pre-Trained Object Detection Models: Performance, Efficiency, and Application Suitabilit
Samain Abid, Muhammad Haris, Mansoor Iqbal, Nadia Khan, Khalid Munir, Hamza Yousaf · 2024
This study conducts a detailed comparative evaluation of pre-trained object detection models in TensorFlow, including SSD, EfficientDet, RetinaNet, Faster R-CNN, and YOLOv4. Performance metrics such as accuracy, inference time, frames per second (FPS), and memory utilization were assessed using the COCO and Pascal VOC datasets. The results indicate that SSD provides a balanced trade-off between speed and accuracy, making it suitable for scenarios requiring moderate performance. EfficientDet demonstrates superior resource efficiency while achieving the highest accuracy, positioning it as the most appropriate model for applications requiring high accuracy and resource optimization. RetinaNet, although slower, excels in detecting small objects, offering advantages in precision for intricate object detection tasks. Faster R-CNN outperforms other models in handling complex, cluttered images but at the cost of real-time performance. YOLOv4, despite challenges in handling overlapping and small objects, delivers the fastest inference time, making it highly effective for real-time applications. This comparative analysis underscores each model's distinct advantages and limitations, providing insights into their suitability based on specific operational requirements.