2021-基于深度学习的人脸检测综述文献-摘要

The Elements of End-to-end Deep Face Recognition: A Survey of Recent Advances
该文献对人脸识别系统中的face detection，face alignment，representation三大模块近年来的成果做了总结，我将综述中关于face detection的重点记录下来，方便自己查阅

1.近年来具有代表性的人脸检测技术时间轴

2.深度人脸检测算法的分类统计.

3.主要的人脸检测数据集总结

对datasets 的总结：
We introduce several widely used datasets for face detection. The statistics of them are given in Table 2. Among them, FDDB [225] is a classic dataset of unconstrained face detection which includes lowresolution, occluded faces and difficult pose variations. It is noteworthy that FDDB uses ellipse as ground-truth instead of rectangular box. The images in PASCAL faces dataset [277] are taken from the Pascal person layout dataset [56]. MALF [278] is designed for finegrained evaluation of face detection in the wild. MAFA [61] is a masked face detection benchmark with various orientations and occlusion degrees. WIDER FACE [285] provides a large number of training data and a challenging test benchmark with large variations.

4. state-of-the-art的方法在 WIDER FACE 上验证集和测试集的表现（指标为AP）

性能比较：
Table 3 shows the performance comparison of the existing face detectors on WIDER FACE validation and test subsets. From the viewpoint of subcategory, we can observe the single-stage methods with anchor-based mechanism (e.g., RefineFace [316], HAMBox [145]) dominate the state-of-the-art performance. For many real-world applications,MTCNN [312], Faceboxes [319], and RetinaFace [40] are the three widely used face detectors for building a face recognition system, since they can achieve well balance on the detection accuracy and efficiency.

5.人脸检测对后续工作的影响
Face detection is the very first procedure in the end-to-end face recognition system, and thereby plays the role of input towards face alignment and face representationThe quality of detection bounding box directly influences on the performance of the subsequent alignment.