<![CDATA[Two-wheeled vehicles dominated the transportation landscape in Indonesia, yet automated monitoring faced challenges due to the complexity of multilabel detection in dense traffic conditions. This research aimed to conduct a comparative analysis between the Detection Transformer (DETR) and You Only Look Once (YOLO) models for multilabel object detection on motorcycles, specifically identifying vehicle types and license plates simultaneously. The research workflow involved multi-label annotating 550 surveillance images, partitioning the dataset using a 70:20:10 hold-out ratio, and applying spatial augmentations (zooming and shearing) to enhance generalization. Subsequently, the DETR, YOLOv11, and YOLOv12 architectures were trained and evaluated across six input resolution scales ranging from 128px to 640px based on mean Average Precision (mAP) metrics and training convergence rates. The results indicated that the DETR model at 640px resolution achieved the highest performance with a mean Average Precision (mAP@50) of 96.97%. Another significant finding revealed that DETR exhibited superior training efficiency compared to the YOLO family, where optimal mAP convergence was reached within only 1 to 7 epochs. In contrast, YOLO models required significantly more training iterations to achieve comparable accuracy levels. This study concluded that the attention mechanism in transformer-based architectures provided higher effectiveness in learning multilabel features for two-wheeled vehicles than conventional convolutional approaches.]]>
