biubug6 / Pytorch_Retinaface

Retinaface get 80.99% in widerface hard val using mobilenet0.25.
MIT License
2.63k stars 774 forks source link

Any difference between this mnet.25 model and that of Face-Detector-1MB-with-landmark? #72

Closed ufoym closed 4 years ago

ufoym commented 4 years ago

I've tested the mnet.25 model shared in https://github.com/biubug6/Face-Detector-1MB-with-landmark. The results are as follow:

The result, when scale = 320px, is consistent with the table you show in https://github.com/biubug6/Face-Detector-1MB-with-landmark (E0.81-M0.71-H0.36 V.S. E0.82-M0.72-H0.37). However, when it comes to the original scale, the result is not that good (E0.76-M0.76-H0.67). But the result shown in the repo is excellent (H0.91-M0.88-H0.74).

So here comes my question: are these pretrained models equivalent? Thanks pretty much!