A Comparative Study of YOLOv5 and YOLOv7 Modifications for Face Detection on a Custom Dataset

Abstract

[eng] Face detection is a fundamental task in computer vision with applications spanning facial recognition, pose estimation, and human-robot interaction. This thesis presents a comprehensive comparative study of two modified versions of the YOLO (You Only Look Once) algorithm, YOLOv5face and YOLOv7face, tailored for landmark detection on a custom dataset of human faces. The study evaluates these models on various aspects, including architecture, accuracy, speed, generalization capability, and specific features. YOLOv5face strikes a balance between accuracy and speed, rendering it suitable for real-time or near-real-time applications. Equipped with a landmark regression head, it excels in tasks requiring precise facial landmark detection. YOLOv7face, on the other hand, outperforms YOLOv5face in accuracy, even in challenging conditions like occlusion and varying lighting. Its robustness positions it as a reliable choice for real-world applications. The comparative analysis underscores the importance of selecting the right model based on specific requirements. YOLOv5face offers efficiency and versatility, while YOLOv7face prioritizes accuracy and robustness. Future research directions include diversifying datasets, fine-tuning, real-world testing, efficiency improvements, and applications in human-robot interaction. This study contributes to the advancement of facial keypoint detection algorithms and guides researchers and practitioners in choosing appropriate models for various computer vision tasks.