CenterFace: Face as Point
实用的边缘设备无锚人脸检测与对齐算法Centerface, 模型大小7.3M。
CenterFace-small 性能达到centerface的同时模型大小仅为2.3M。
![image](results/bl4.jpg)
- OpenCV 4.1.0
- Numpy
- Python3.6+
Model Version |
Easy Set |
Medium Set |
Hard Set |
FaceBoxes |
0.840 |
0.766 |
0.395 |
FaceBoxes3.2× |
0.798 |
0.802 |
0.715 |
RetinaFace-mnet |
0.887 |
0.870 |
0.792 |
LFFD-v1 |
0.910 |
0.881 |
0.780 |
LFFD-v2 |
0.837 |
0.835 |
0.729 |
CenterFace |
0.935 |
0.924 |
0.875 |
CenterFace-small |
0.931 |
0.924 |
0.870 |
Model Version |
Easy Set |
Medium Set |
Hard Set |
FaceBoxes |
0.839 |
0.763 |
0.396 |
FaceBoxes3.2× |
0.791 |
0.794 |
0.715 |
LFFD-v1 |
0.910 |
0.881 |
0.780 |
LFFD-v2 |
0.837 |
0.835 |
0.729 |
CenterFace |
0.932 |
0.921 |
0.873 |
- RetinaFace-mnet 是RetinaFace-MobileNet-0.25的简写,来自于非常好的工作insightface。
- LFFD-v1 也是很好的工作LFFD。
- CenterFace/CenterFace-small的测试方法是MULTI-SCALE,因为训练图像和测试图像尺度的不一致性,多尺度测试才能反应centerface的真实性能。
不过,对于SIO(原图单次推理),CenterFace在val集上也可以达到:92.2% (Easy), 91.1% (Medium) and 78.2% 。
而RetinaFace-mnet在val集上是:89.6% (Easy), 87.1% (Medium) and 68.1%
Model Version |
Disc ROC curves score |
RetinaFace-mnet |
96.0@1000 |
LFFD-v1 |
97.3@1000 |
LFFD-v2 |
97.2@1000 |
CenterFace |
98.0@1000 |
CenterFace-small |
98.1@1000 |
Resolution-> |
640×480 |
1280×720(704) |
1920×1080(1056) |
RetinaFace-mnet |
5.40ms |
6.31ms |
10.26ms |
LFFD-v1 |
7.24ms |
14.58ms |
28.36ms |
CenterFace |
5.5ms |
6.4ms |
8.7ms |
CenterFace-small |
4.4ms |
5.7ms |
7.3ms |
![image](results/box_lm.jpg)
![image](results/bl3.jpg)
![image](results/bl1.jpg)
欢迎加入 QQ Group(912759877) 交流讨论, 包括但不限:人脸检测、稠密对齐、活体、3D重建等。