【YOLO5Face】《YOLO5Face：Why Reinventing a Face Detector》

arXiv-2021

---

---

1 Background and Motivation

人脸检测是非常基础的 CV 应用之一，作为人脸识别、验证、跟踪、对齐、表情分析等诸多任务的第一步，吸引了学术界和业界的许多研究和发展

作者 treat face detection as a general object detection task（face detection is just a sub task of general object detection），在 yolov5 目标检测的工程基础上，改进提出人脸检测器 YOLO5Face

2 Related Work

• Object Detection

  • traditional
  • two-stage
  • one-stage

• Face Detection
  解决 scale, pose, occlusion, expression, makeup, illumination, blur and etc 问题

• YOLO

3 Advantages / Contributions

• 设计 YOLO5Face 人脸检测器
• 针对不用应用需求，提出不同大小的人脸检测器
• 在 WiderFace 上评估，实现了 SOTA（validation）

4 Method

1）Network Architecture

相比于 yolov5 的改进

• add a landmark regression head to the YOLOv5 network. 采用的是 Wing Loss
• replace the Focus with a Stem block structure（图 1 的（d））
• change the SPP block and use a smaller kernel（13x13-9x9-5x5 改成了 7x7-5x7-3x3）
• add a P6 output block with stride of 64（增加大脸检出率）
• 调增了 DA 策略，取消了上下翻转，Mosaic 和小目标兼容性不好，random cropping 效果不错
• design two super light-weight models based on ShuffleNetV2（YOLOv5n / YOLOv5n-0.5）

2）Landmark Regression

L1 L2 和 smooth L1，these loss functions are not sensitive to small errors.

采用的 wing loss 对小 error 更敏感


最终的 loss 由 objection loss 和 landmark loss 构成

5 Experiments

5.1 Datasets

• WiderFace
  • contains 32,203 images and 393,703 faces
  • train/validation/test sets by ratio 50%/10%/40%
  • three levels of difficulty: Easy, Medium, and Hard.
• FDDB
  • 5171 faces annotated in 2845 images.

5.2 Ablation Study

模型结构细节

消融实验

• Stem Block vs. Focus Layer
• SPP with Smaller Size Kernels
• P6 Output Block
• Data Augmentation
  • Mosaic helps the mAP in the Hard dataset.
  • the Mosaic has to work with the ignoring small faces, otherwise the performance degrades dramatically

5.3 YOLO5Face for Face Recognition

关键点的对比，vs RetinaFace


大角度作者的方法会更准确一点

5.4 YOLO5Face on WiderFace Dataset

SCRFD 感觉好猛

看看 PR 曲线

validation dataset：YOLOv5x6-Face detector achieves 96.9%, 96.0%, 91.6% mAP on the Easy, Medium, and Hard subset（比SOTA 猛）

test dataset：YOLOv5x6-Face detector achieves 95.8%, 94.9%, 90.5% mAP on the Easy, Medium, and Hard subset（没 SOTA猛）

we only use multiple scales and left-right flipping without using other test-time augmentation (TTA) methods.

5.5 YOLO5Face on FDDB Dataset

6 Conclusion（own） / Future work

• VGA resolution input images