LFFD: A Light and Fast Face Detector for Edge Devices

paper link

Requirements:

• pytorch>=1.2
• opencv-python
• torchvision>=0.4

Label Assignment

Boxes are in different scale. Anchor points with more than one matched gt box are ignored.
green box --> gt box
blue/red point --> positive sample
The positive sample sampling of the fifth branch is shown below.

Different settings from the paper

• regression target
• add BN
• more classes

Test

I have trained the model on a safety helmet dataset without any augmentation. It seems to be working well.

tensorboard

test_result