Improving Vehicle Detection in Challenging Datasets: YOLOv5s and Frozen Layers Analysis

Abstract

Small datasets and imbalanced classes often cause problems when it used as primary research material. In case of classification and object detection, some researchers proposed Transfer Learning (TF) with several frozen layers. Moreover, YOLO (You Only Look Once) is one of the algorithms that works in real-time object detection. In this research, we focused on evaluating the YOLOv5s version of detecting vehicles in small and imbalanced datasets. The original YOLOv5s were trained and compared with YOLOv5s with freezing layers method (10 and 24 frozen layers). The experimental results of original YOLOv5s were precision score of 0.779, recall value of 0.933, mAP@0.5 of 0.93 and mAP@0.5:0.95 of 0.684 while YOLOv5s with 10 frozen layers where precision score was decreased to 0.639, but the other value increase with recall value of 0.939, mAP@0.5 of 0.951 and mAP@0.5:0.95 of 0.732.  Overall, the version with 10 frozen layers demonstrated superior performance in addressing the challenges of small and imbalanced datasets, particularly excelling in recall and mAP metrics.