about the error recognition

[leavessnow26]

Thanks for your share, and it's a good job. However when I test the trump.mp4 with the supplied mlp.pkl, I found some error results and the prob is very high. Is there something wrong with me? or how can solve it?

[1996scarlet]

TL;DR follow this link

So maybe you could retrain the mlp.pkl with your own data(using train_mlp.py), or you can try to use others classifier(even the softmax), or fine-tuning.

By the way, the ArcFace Paper said they used:

• model ensemble (vote)
• threelayer fully connected network (MLP), my mlp just have one layer (lol)

[ThanhPham1987]

Hi @leavessnow26 Would you like to share your experience to run usb_camera.py file? Please help me! Best regards, PeterPham

[brealisty]

have u ever try to another way to indetify face, for example: after get the 512-D feature, then compute the distance withe face database. compare with a classifer, which way is better? I just know the distance way, is easy to expand face(number of class), but I don't know which way can get a higher acc.