Perbandingan Kinerja YOLOv8 dan SSD MobileNet untuk Deteksi Kelengkapan Atribut Seragam Siswa

Abstract

Inspecting the completeness of student uniform attributes is challenging because several items, such
as epaulettes, name tags, belts, black shoes, and caps, may appear small, partially occluded, or
visually unclear. This study compares YOLOv8n and SSD MobileNet for detecting student uniform
attributes using a self-collected dataset. The dataset contains 1,532 augmented images, five object
classes, and 5,294 YOLO-format bounding box annotations. The experimental procedure included
dataset preparation, object annotation, train-validation-test splitting, model training, and final
evaluation on the test set. The evaluation used precision, recall, mAP@50, mAP@50-95, inference
time, and FPS. On the test set, YOLOv8n achieved a precision of 0.8966, recall of 0.8689, mAP@50
of 0.9276, mAP@50-95 of 0.6723, inference time of 10.36 ms, and 96.47 FPS. SSD MobileNet
achieved a precision of 0.5140, recall of 0.5169, mAP@50 of 0.6034, mAP@50-95 of 0.3171,
inference time of 11.00 ms, and 90.87 FPS. These findings indicate that YOLOv8n performed better
in the tested dataset and experimental setting.