Object Detection for Vietnamese Traffic Signs
Faster R-CNN, YOLOv12, FCOS, RetinaNet, and SSD300
A comparison of one-stage and two-stage detectors fine-tuned on the Vietnamese Traffic Signs dataset. Traffic signs are a demanding target: the objects are small relative to the frame, and several classes differ only in their pictogram.
Models. Pre-trained Faster R-CNN, YOLOv12, FCOS, RetinaNet, and SSD300, each fine-tuned on the same data for a like-for-like comparison.
Dataset. Bounding boxes and class targets were labeled by hand for signage in the Ho Chi Minh City context, and the annotated dataset was published on Kaggle.