1. What I want to Know
How they add key point landmark detection
How they design their network due to this change
2. Yolo5Face Contribution
- Redesign the YoloV5
- Design models for different model size
- Evaluate on Widerface
3. Key Modification
Add a landmark regression head, this is what I want to know
In face detection, there are 5 points (simplified from 68 points)
- General loss functions for landmark regression: Wing-loss
$$
\operatorname{wing}(x)= \begin{cases}w \cdot \ln (1+|x| / e), & \text { if } x<w \ |x|-C, & \text { otherwise }\end{cases}
$$
- Add landmark loss with object detection loss
$$
loss(s)=loss_{obj}+\lambda{L} loss_L
$$
Stem Block
- YOLOv5 has a focus layer, replacing focus layer with stem block
- Adding landmark into the stem block