<![CDATA[Face detection in dense and occluded environments remains a significant challenge in computer vision. This study compares the CNN-based YOLOv12 and the Transformer-based RF-DETR to determine the optimal balance between accuracy and latency for resource-constrained edge computing. Using the WIDER FACE dataset and an NVIDIA T4 GPU, multiple model variants were evaluated. Due to GPU memory constraints during training of the RF-DETR Medium variant, a standardized batch size of 8 was implemented across all models. To ensure methodological rigor, quantitative metrics (precision, recall, F1-score, mAP) were strictly assessed on the validation set. Concurrently, a 100-image subset of the test set was used exclusively for inference efficiency benchmarking, completely separate from detection evaluation. Results indicate YOLOv12X achieved superior overall detection performance (F1-score: 0.764, mAP@50:95: 0.440), significantly outperforming RF-DETR Medium. For real-time applications, YOLOv12M demonstrated the highest efficiency (36.17 FPS vs. 23.32 FPS). Qualitatively, YOLOv12 maintained high sensitivity in crowded scenes, whereas RF-DETR provided stable small-scale face detection despite its lower recall. Overall, under these constrained-hardware conditions, YOLOv12 appears to be a highly viable solution for surveillance systems, while RF-DETR offers a stable alternative for small-object detection when computational overhead and training budgets are less restrictive.]]>
