Key points are not available for this paper at this time.
Face detection is a primary and essential task in the domain of face recognition and identification, serving as the initial and crucial step in the process. There are several models for face detection, such as RetinaFace, Yolov5-Face, and SCRFD. The best model can vary based on specific criteria such as accuracy, speed, memory efficiency, and suitability for different applications. Among these face detection models, the SCRFD model has the most optimal trade-off between speed and accuracy, which makes it a popular and suitable model for face detection applications. However, in the real world, extracting meaningful features and detecting small faces in surveillance videos can be challenging due to the low resolution. In this research, we focus on enhancing the accuracy of the SCRFD(Sample and Computation Redistribution for Efficient Face Detection) model on small faces. We conduct a comprehensive enhancement using multiple strategies, including expanding the dataset, adjusting input shape during training time, tuning weight decay, optimizing the learning rate scheduler, and testing different augmentations. By analyzing these factors systematically, we achieved mAP of 69.88% and 85.23% at 640 × 640 and 1280 × 1280 resolutions respectively in the Hard dataset, which outperforms the original SCRFD by 0.39% and 1.58%.
Ebrahimian et al. (Tue,) studied this question.