YOLOv3: Face Detection in Complex Environments

Face detection has been well studied for many years. However, the problem of face detection in complex environments is still being studied. In complex environments, faces is often blocked and blurred. This article proposes applying YOLOv3 to face detection problems in complex environments. First, we will re-cluster the data set in order to find the most suitable a priori box. Then we set multiple score values to make it possible to predict the results of multiple sets of images and find the optimal score value. Experimental results show that after adjustment, the model has more advantages in face detection than the original model in complex environments. The average accuracy is more than 10% higher than that of aggregate channel feature (ACF), Towstage convolutional neural network (CNN) and multi-scale Cascade CNN in face detection benchmarks WIDER FACE.