PROJECT TITLE :

Anchor Cascade for Efficient Face Detection

ABSTRACT:

To perform facial analysis tasks like facial reenactment and face identification, face detection is needed. These two face detectors have gained a lot of attention from the community because of their impressive demos. In contrast, anchor-based face detectors require huge Neural Networks pre-trained on Image Classification datasets like ImageNet, which are computationally inefficient for both training and deployment of the face detectors. However, cascade face detectors frequently suffer from low detection accuracy. The anchor cascade system we present in this paper is an effective anchor-based cascade structure. Further, we suggest a context pyramid maxout mechanism for anchor cascade in order to improve the detection accuracy. The high accuracy and efficiency of face detection models can be trained via anchor cascade. Anchor cascade face detector considerably improves detection accuracy, for example, from 0.9435 to 0.9704 at FDDB false positive rate of 1k false positives compared to the popular Convolutional Neural Network (CNN)-based cascade face detector MTCNN. The suggested framework has been tested on two widely used face detection benchmarks: FDDB and WIDER FACE.