Object Detection Model Performance Analysis for Garbage Detection on Resource-Limited Platforms

Object detection on devices with limited computational resources presents a formidable challenge due to the intricate balance between accuracy and efficiency. This paper investigates the performance of three leading object detection models—MobileNet SSD v2, YOLOv8 Nano, and EfficientDet—trained on a garbage dataset encompassing a single class. Evaluation of these models is conducted using four fundamental metrics—precision, recall, mean average precision (mAP), and F1 score. Notably, MobileNet SSD v2 demonstrates outstanding performance, achieving a precision of 0.9913. Conversely, EfficientDet exhibits subpar performance despite a precision score of 1.0, attributed to a significantly lower recall of 0.16. The models are deployed on an overclocked Raspberry Pi equipped with a camera module for real-time garbage detection. Analysis of the study's results, including confusion matrices, detection images, frames per second (FPS), and confidence scores, provides insights into the ornate trade-offs between model complexity, accuracy, and computational efficiency in the context of garbage detection on low-power devices.