Pre-trained face detection models such as MTCNN achieve strong results on standard benchmarks, however, it remains unclear how their performance holds up under varying real world conditions. This paper evaluates MTCNN on the WiderFace dataset across three difficulty levels of images, finding that precision remains stable, dropping from 0.93 to 0.88, while recall declines sharply from 0.94 to 0.33. This degradation is explained by a larger presence of small, overlapping faces in crowded scenes as image complexity increases. A transfer learning experiment was conducted using MobileNetV2 to explore whether a general purpose pre-trained model can be adapted to a specific scene classification task, which scored 68% for test accuracy. An interesting methodological finding revealed that whole images substantially outperformed cropped faces as model input, due to the loss of contextual cues and environmental details that cropping caused. Data augmentation, similarly, was found to decrease performance on the small dataset instead of improving it. Cross demographic robustness experiments revealed that models trained on one demographic group generalise poorly to others, emphasizing the importance of both data diversity and data size in robustness of a model. Finally, inference and training speed are measured and compared between CPU and GPU on Apple’s M1 hardware, showing comparable performance. Collectively these findings illustrate the limitations of deploying pre-trained face detection models under real world conditions and the trade-offs involved in adapting them through transfer learning.
Additional Key Words and Phrases: face detection, MTCNN, WiderFace, transfer learning, MobileNetV2, CPU/GPU performance, model robustness