YongtaoGe / RetinaFace

Reimplement RetinaFace using PyTorch.

RetinaFace.PyTorch

Reimplentment of RetinaFace with modification!

Requirements

• pytorch
• tvm

WIDER accuracy

Impelmentation	Easy-Set	Medium-Set	Hard-Set	Speed	Test-Method(short side 1600)	Params	FLOPs (640*640)
res18	-	-	-	-	single-scale	-
modified_mnasnet	0.8495	0.8339	0.7531	-	single-scale	325.38k	386.57MMac
modified_mnasnet	0.8649	0.8436	0.7139		single-scale	325.38k	386.57MMac
mobilenet0.25	-	-	-	-	single-scale	427.22k	1.02G

Note: I test the forward speed with converted tvm model using mac i7 cpu.

AFLW accuracy

demo:

TODO:

• Wider Face accuracy testing
• AFLW accuracy testing
• Inference on TVM