How exactly were the pretrained models trained?

[tobiasbrechtelsbauer]

Hi kensun0,

thank you for this awesome project! I am currently using it for the practical part of my masters thesis in information technology. You and this github repository will be cited there! 🥇
This is also no real issue, the code is working well. I just need some information and because you set your email to noreply@github I couldn't contact you by email.

Could you please tell me how the pretrained models were trained? Well I can see the number of iterations in the filenames and also that OHEM was used. I just guess they were trained with the WIDER Face train dataset. Or did you enlarge the training data with some other datasets?
If you have more detailed information about what parameters (learning-rate for example) you used to create the models I would love know them.

Thank you very much for this project and also thank you very much in advance for answering my questions!

Greetings
Tobias

[kensun0]

Hi tobiasbrechtelsbauer,

I get the impression that i use ImageNet to pretrain the model, then fixxed the parameters of layers which are front of RPN layer.
I only use WIDER to train the uploaded model with OHEM. Data augument operations were used, for example, flip, color jitter, random crop, resize, smooth, sharp, etc.