CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point

概述 anchor free的人脸检测方法,同时定位5个关键点。输入图为640X480时,cpu可达到30FPS,性能也不错。

WIDER FACE Val/Test-Easy: 0.935/0.932, Medium: 0.924/0.921, Hard: 0.875/0.873 and FDDB discontinuous: 0.980, continuous: 0.732
  • 通过热图预测人脸中心
  • 对每一个人脸中心预测一个box,offset,5个landmarks
  • 主干网络为mobilenet2+FPN
  • 输入图保证长宽为32的整数倍,输出图相对输入stride为4
  • paper
  • github:只提供了demo和model,无训练代码。
  • 实际测试在自拍大头照上比较差。可能跟训练样本有关。
细节
  • 模型输出为:heatmap(B,1,H/4,W/4),scale(B,2,H/4,W/4),offset(B,2,H/4,W/4),landmark(B,10,H/4,W/4)
  • heatmap分支同objects as points,采用focal loss
    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    focal loss.png
  • 由于热图只能预测整数位置,故加上偏移
    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    Loff.png

    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    off.png
  • scale
    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    size.png
  • landmarks

    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    landmark.png
  • 【CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point】总体loss

    CenterFace:|CenterFace: Joint Face Detection and Alignment Using Face as Point
    文章图片
    total loss.png
  • 增广:随机反转,随机scale,color jitter,随机crop方块resize到800*800

    推荐阅读