Comparison Of Efficientnet And Yolov8 Algorithms In Motor Vehicle Classification

Abstract

The YOLOv8 accuracy curve highlights clear overfitting. As shown in the graph, the model reaches 100% training accuracy from the first epoch and remains flat, indicating it memorized the training data. However, validation accuracy lags behind, fluctuating between 90% and 92% without significant improvement. This discrepancy between training and validation performance suggests that YOLOv8 struggles to generalize to unseen data. The issue likely stems from its architecture, which is optimized for object detection tasks that prioritize object localization over feature extraction for classification. When repurposed for classification, YOLOv8 may not extract the nuanced visual patterns needed to differentiate similar classes, such as trucks and buses. Consequently, although YOLOv8 performs well on the training set, its classification accuracy in real-world scenarios is limited. Addressing this may require architectural adjustments, stronger regularization, or more diverse training data to enhance the model’s generalization for pure classification tasks.